研究者詳細

顔写真

オカタニ タカユキ
岡谷 貴之
Takayuki Okatani
所属
大学院情報科学研究科 システム情報科学専攻 知能ロボティクス学講座(イメージ解析学分野)
職名
教授
学位
  • 博士(工学)(東京大学)

  • 修士(工学)(東京大学)

経歴 2

  • 2016年10月 ~ 継続中
    理化学研究所革新知能統合研究センター

  • 2003年11月 ~ 2004年10月
    Rutgers大学先進情報処理センター

学歴 2

  • 東京大学 工学系研究科 計数工学

    ~ 1999年3月

  • 東京大学 工学部 計数工

    ~ 1994年3月

所属学協会 4

  • 計測自動制御学会

  • 電子情報通信学会

  • 情報処理学会

  • IEEE

研究キーワード 6

  • マルチモーダルAI

  • 多視点幾何

  • 画像理解

  • 画像認識

  • 深層学習

  • コンピュータビジョン

研究分野 2

  • 情報通信 / 知能ロボティクス / コンピュータビジョン

  • 情報通信 / 知覚情報処理 / コンピュータビジョン

受賞 11

  1. 画像の認識・理解シンポジウム(MIRU)2017優秀論文賞

    2017年8月 画像の認識・理解シンポジウム実行委員会

  2. Best paper honorable mention

    2016年10月30日 PACLIC2016 award committee Recognizing Open-Vocabulary Relations between Objects in Images

  3. MIRU優秀発表賞

    2016年8月3日 MIRU実行委員会 Learning to Describe Product Images from Noisy Online Data

  4. MIRU優秀発表賞

    2015年8月1日 MIRU運営委員会 Joint Clothing Detection

  5. MIRUフロンティア賞

    2015年8月1日 MIRU運営委員会 Scene Change Detection from a Pair of Roughly Registered Images Based on CNN Features and Superpixel Segmentation

  6. Sang Uk Lee Best Student Paper Award

    2014年12月1日 ACCV award committee Separation of Reflection Components by Sparse Non-negative Matrix Factorization

  7. Best Application Paper Honorable Mention

    2014年12月1日 ACCV award committee Massive City-scale Surface Condition Analysis using Ground and Aerial Imagery

  8. 画像の認識理解シンポジウム(MIRU2010) MIRU長尾賞

    2010年7月 画像の認識・理解シンポジウム実行委員会

  9. 画像の認識理解シンポジウム(MIRU2007)優秀論文賞

    2007年7月31日 電子情報通信学会,情報処理学会他

  10. 画像の認識理解シンポジウム(MIRU2006)優秀論文賞

    2006年7月20日 電子情報通信学会,情報処理学会他

  11. 情報処理学会平成9年度山下記念研究賞

    1997年1月1日 情報処理学会

︎全件表示 ︎最初の5件までを表示

論文 175

  1. Inverting the Generation Process of Denoising Diffusion Implicit Models: Empirical Evaluation and a Novel Method.

    Yan Zeng, Masanori Suganuma, Takayuki Okatani

    WACV 4516-4524 2025年

    DOI: 10.1109/WACV61041.2025.00453  

  2. RefVSR++: Exploiting Reference Inputs for Reference-based Video Super-resolution.

    Han Zou, Masanori Suganuma, Takayuki Okatani

    WACV 2756-2765 2025年

    DOI: 10.1109/WACV61041.2025.00273  

  3. Deep learning-based bridge damage cause estimation from multiple images using visual question answering

    Tatsuro Yamane, Pang-jo Chun, Ji Dang, Takayuki Okatani

    Structure and Infrastructure Engineering 1-14 2024年5月20日

    出版者・発行元: Informa UK Limited

    DOI: 10.1080/15732479.2024.2355929  

    ISSN:1573-2479

    eISSN:1744-8980

  4. Symmetry-aware Neural Architecture for Embodied Visual Navigation.

    Shuang Liu, Masanori Suganuma, Takayuki Okatani

    International Journal of Computer Vision 132 (4) 1091-1107 2024年4月

    DOI: 10.1007/s11263-023-01909-4  

  5. That's BAD: blind anomaly detection by implicit local feature clustering.

    Jie Zhang, Masanori Suganuma, Takayuki Okatani

    Machine Vision and Applications 35 (2) 31-31 2024年3月

    DOI: 10.1007/s00138-024-01511-9  

  6. Improving visual question answering for bridge inspection by pre-training with external data of image-text pairs.

    Thannarot Kunlamai, Tatsuro Yamane, Masanori Suganuma, Pang-jo Chun, Takayuki Okatani

    Computer-Aided Civil and Infrastructure Engineering 39 (3) 345-361 2024年2月

    DOI: 10.1111/mice.13086  

  7. Action-Agnostic Point-Level Supervision for Temporal Action Detection.

    Shuhei M. Yoshida, Takashi Shibata 0001, Makoto Terao, Takayuki Okatani, Masashi Sugiyama

    CoRR abs/2412.21205 2024年

    DOI: 10.48550/arXiv.2412.21205  

  8. Exploring the Potential of Multi-Modal AI for Driving Hazard Prediction

    Korawat Charoenpitaks, Van Quang Nguyen, Masanori Suganuma, Masahiro Takahashi, Ryoma Niihara, Takayuki Okatani

    IEEE Transactions on Intelligent Vehicles 2024年

    DOI: 10.1109/TIV.2024.3417353  

    eISSN:2379-8858

  9. Appearance-Based Curriculum for Semi-Supervised Learning with Multi-Angle Unlabeled Data.

    Yuki Tanaka, Shuhei M. Yoshida, Takashi Shibata, Makoto Terao, Takayuki Okatani, Masashi Sugiyama

    IEEE/CVF Winter Conference on Applications of Computer Vision(WACV) 2768-2777 2024年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV57701.2024.00276  

  10. SBCFormer: Lightweight Network Capable of Full-size ImageNet Classification at 1 FPS on Single Board Computers.

    Xiangyong Lu, Masanori Suganuma, Takayuki Okatani

    IEEE/CVF Winter Conference on Applications of Computer Vision(WACV) 1112-1122 2024年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV57701.2024.00116  

  11. Contextual Affinity Distillation for Image Anomaly Detection.

    Jie Zhang, Masanori Suganuma, Takayuki Okatani

    IEEE/CVF Winter Conference on Applications of Computer Vision(WACV) 148-157 2024年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV57701.2024.00022  

  12. Globalizing Local Features: Image Retrieval Using Shared Local Features with Pose Estimation for Faster Visual Localization.

    Wenzheng Song, Ran Yan, Boshu Lei, Takayuki Okatani

    IEEE International Conference on Robotics and Automation(ICRA) 6290-6297 2024年

    出版者・発行元: IEEE

    DOI: 10.1109/ICRA57147.2024.10610786  

  13. Temporal Insight Enhancement: Mitigating Temporal Hallucination in Video Understanding by Multimodal Large Language Models.

    Li Sun 0007, Liuan Wang, Jun Sun 0004, Takayuki Okatani

    ICPR (7) 455-473 2024年

    DOI: 10.1007/978-3-031-78183-4_29  

  14. Rethinking unsupervised domain adaptation for semantic segmentation.

    Zhijie Wang, Masanori Suganuma, Takayuki Okatani

    Pattern Recognition Letters 186 119-125 2024年

    DOI: 10.1016/j.patrec.2024.09.022  

  15. Improved high dynamic range imaging using multi-scale feature flows balanced between task-orientedness and accuracy.

    Qian Ye, Masanori Suganuma, Takayuki Okatani

    Computer Vision and Image Understanding 248 104126-104126 2024年

    DOI: 10.1016/j.cviu.2024.104126  

  16. Unsupervised domain adaptation for semantic segmentation via cross-region alignment

    Zhijie Wang, Xing Liu, Masanori Suganuma, Takayuki Okatani

    Computer Vision and Image Understanding 234 2023年9月

    DOI: 10.1016/j.cviu.2023.103743  

    ISSN:1077-3142

    eISSN:1090-235X

  17. Zero-shot temporal event localisation: Label-free, training-free, domain-free

    Li Sun, Ping Wang, Liuan Wang, Jun Sun, Takayuki Okatani

    IET Computer Vision 17 (5) 599-613 2023年8月

    DOI: 10.1049/cvi2.12224  

    ISSN:1751-9632

    eISSN:1751-9640

  18. B-CNN: a deep learning method for accelerometer-based fatigue cracks monitoring system

    Yanjie Zhu, Hidehiko Sekiya, Takayuki Okatani, Masayuki Tai, Shogo Morichika

    Journal of Civil Structural Health Monitoring 13 (4-5) 947-959 2023年3月31日

    出版者・発行元: Springer Science and Business Media LLC

    DOI: 10.1007/s13349-023-00690-9  

    ISSN:2190-5452

    eISSN:2190-5479

  19. Zero-shot versus Many-shot: Unsupervised Texture Anomaly Detection

    Toshimichi Aota, Lloyd Teh Tzer Tong, Takayuki Okatani

    Proceedings - 2023 IEEE Winter Conference on Applications of Computer Vision, WACV 2023 5553-5561 2023年

    DOI: 10.1109/WACV56688.2023.00552  

  20. Bright as the Sun: In-depth Analysis of Imagination-Driven Image Captioning

    Huyen Thi Thanh Tran, Takayuki Okatani

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 13844 LNCS 675-691 2023年

    DOI: 10.1007/978-3-031-26316-3_40  

    ISSN:0302-9743

    eISSN:1611-3349

  21. Improving visual question answering for bridge inspection by pre-training with external data of image–text pairs

    Thannarot Kunlamai, Tatsuro Yamane, Masanori Suganuma, Pang Jo Chun, Takayaki Okatani

    Computer-Aided Civil and Infrastructure Engineering 2023年

    DOI: 10.1111/mice.13086  

    ISSN:1093-9687

    eISSN:1467-8667

  22. How Do Label Errors Affect Thin Crack Detection by DNNs

    Liang Xu, Han Zou, Takayuki Okatani

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops 2023-June 4414-4423 2023年

    DOI: 10.1109/CVPRW59228.2023.00464  

    ISSN:2160-7508

    eISSN:2160-7516

  23. Geometry Enhanced Reference-based Image Super-resolution

    Han Zou, Liang Xu, Takayuki Okatani

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops 2023-June 6124-6133 2023年

    DOI: 10.1109/CVPRW59228.2023.00652  

    ISSN:2160-7508

    eISSN:2160-7516

  24. Network Pruning and Fine-tuning for Few-shot Industrial Image Anomaly Detection

    Jie Zhang, Masanori Suganuma, Takayuki Okatani

    IEEE International Conference on Industrial Informatics (INDIN) 2023-July 2023年

    DOI: 10.1109/INDIN51400.2023.10218283  

    ISSN:1935-4576

  25. Prompt Prototype Learning Based on Ranking Instruction For Few-Shot Visual Tasks

    Li Sun, Liuan Wang, Jun Sun, Takayuki Okatani

    Proceedings - International Conference on Image Processing, ICIP 3235-3239 2023年

    DOI: 10.1109/ICIP49359.2023.10222039  

    ISSN:1522-4880

  26. Accurate Single-Image Defocus Deblurring Based on Improved Integration with Defocus Map Estimation

    Qian Ye, Masanori Suganuma, Takayuki Okatani

    Proceedings - International Conference on Image Processing, ICIP 750-754 2023年

    DOI: 10.1109/ICIP49359.2023.10223146  

    ISSN:1522-4880

  27. Bridge Damage Cause Estimation Using Multiple Images Based on Visual Question Answering.

    Tatsuro Yamane, Pang-jo Chun, Ji Dang, Takayuki Okatani

    CoRR abs/2302.09208 2023年

    DOI: 10.48550/arXiv.2302.09208  

  28. A lightweight deep learning model for automatic segmentation and analysis of ophthalmic images

    Parmanand Sharma, Takahiro Ninomiya, Kazuko Omodaka, Naoki Takahashi, Takehiro Miya, Noriko Himori, Takayuki Okatani, Toru Nakazawa

    Scientific Reports 12 (1) 2022年12月

    出版者・発行元: Springer Science and Business Media LLC

    DOI: 10.1038/s41598-022-12486-w  

    eISSN:2045-2322

    詳細を見る 詳細を閉じる

    Abstract Detection, diagnosis, and treatment of ophthalmic diseases depend on extraction of information (features and/or their dimensions) from the images. Deep learning (DL) model are crucial for the automation of it. Here, we report on the development of a lightweight DL model, which can precisely segment/detect the required features automatically. The model utilizes dimensionality reduction of image to extract important features, and channel contraction to allow only the required high-level features necessary for reconstruction of segmented feature image. Performance of present model in detection of glaucoma from optical coherence tomography angiography (OCTA) images of retina is high (area under the receiver-operator characteristic curve AUC ~ 0.81). Bland–Altman analysis gave exceptionally low bias (~ 0.00185), and high Pearson’s correlation coefficient (p = 0.9969) between the parameters determined from manual and DL based segmentation. On the same dataset, bias is an order of magnitude higher (~ 0.0694, p = 0.8534) for commercial software. Present model is 10 times lighter than Unet (popular for biomedical image segmentation) and have a better segmentation accuracy and model training reproducibility (based on the analysis of 3670 OCTA images). High dice similarity coefficient (D) for variety of ophthalmic images suggested it’s wider scope in precise segmentation of images even from other fields. Our concept of channel narrowing is not only important for the segmentation problems, but it can also reduce number of parameters significantly in object classification models. Enhanced disease diagnostic accuracy can be achieved for the resource limited devices (such as mobile phone, Nvidia’s Jetson, Raspberry pi) used in self-monitoring, and tele-screening (memory size of trained model ~ 35 MB).

  29. Development of a continuum robot enhanced with distributed sensors for search and rescue

    Yu Yamauchi, Yuichi Ambe, Hikaru Nagano, Masashi Konyo, Yoshiaki Bando, Eisuke Ito, Solvi Arnold, Kimitoshi Yamazaki, Katsutoshi Itoyama, Takayuki Okatani, Hiroshi G. Okuno, Satoshi Tadokoro

    ROBOMECH Journal 9 (1) 2022年12月

    出版者・発行元: Springer Science and Business Media LLC

    DOI: 10.1186/s40648-022-00223-x  

    eISSN:2197-4225

    詳細を見る 詳細を閉じる

    Abstract Continuum robots can enter narrow spaces and are useful for search and rescue missions in disaster sites. The exploration efficiency at disaster sites improves if the robots can simultaneously acquire several pieces of information. However, a continuum robot that can simultaneously acquire information to such an extent has not yet been designed. This is because attaching multiple sensors to the robot without compromising its body flexibility is challenging. In this study, we installed multiple small sensors in a distributed manner to develop a continuum-robot system with multiple information-gathering functions. In addition, a field experiment with the robot demonstrated that the gathered multiple information has a potential to improve the searching efficiency. Concretely, we developed an active scope camera with sensory functions, which was equipped with a total of 80 distributed sensors, such as inertial measurement units, microphones, speakers, and vibration sensors. Herein, we consider space-saving, noise reduction, and the ease of maintenance for designing the robot. The developed robot can communicate with all the attached sensors even if it is bent with a minimum bending radius of 250 mm. We also developed an operation interface that integrates search-support technologies using the information gathered via sensors. We demonstrated the survivor search procedure in a simulated rubble environment of the Fukushima Robot Test Field. We confirmed that the information provided through the operation interface is useful for searching and finding survivors. The limitations of the designed system are also discussed. The development of such a continuum robot system, with a great potential for several applications, extends the application of continuum robots to disaster management and will benefit the community at large.

  30. GRIT: Faster and Better Image Captioning Transformer Using Dual Visual Features.

    Van-Quang Nguyen, Masanori Suganuma, Takayuki Okatani

    European Conference on Computer Vision (ECCV) 36 167-184 2022年9月

    DOI: 10.1007/978-3-031-20059-5_10  

  31. Real-time vehicle identification using two-step LSTM method for acceleration-based bridge weigh-in-motion system

    Yanjie Zhu, Hidehiko Sekiya, Takayuki Okatani, Ikumasa Yoshida, Shuichi Hirano

    JOURNAL OF CIVIL STRUCTURAL HEALTH MONITORING 12 (3) 689-703 2022年6月

    DOI: 10.1007/s13349-022-00576-2  

    ISSN:2190-5452

    eISSN:2190-5479

  32. Symmetry-aware Neural Architecture for Embodied Visual Exploration.

    Shuang Liu 0002, Takayuki Okatani

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 17221-17230 2022年6月

    DOI: 10.1109/CVPR52688.2022.01673  

    ISSN:0920-5691

    eISSN:1573-1405

  33. 自己教師あり学習を用いた熱画像の解析によるコンクリートの浮き・剥離の自動検出

    川野輪 壮太, 林 詳悟, 岡谷 貴之, Kang-Jun Liu, 全 邦釘

    AI・データサイエンス論文集 3 (J2) 47-55 2022年

    出版者・発行元: 公益社団法人 土木学会

    DOI: 10.11532/jsceiii.3.j2_47  

    eISSN:2435-9262

    詳細を見る 詳細を閉じる

    コンクリートの熱画像を撮影することで内部損傷を遠隔で検出する赤外線サーモグラフィ法は,人間の 判断では損傷部の見落としが多い.また,赤外線法に Convolutional Neural Network(CNN)による自動検 出を導入する動きはすでに始まっているが,教師データが不足しているため十分な精度は得られていない.そこで,本研究では自己教師あり学習に着目した.自己教師あり学習には,教師ラベルが少なくても高い 精度を実現できる可能性がある.本研究では,自己教師あり学習を赤外線法に導入する方法の一例を示し,その有効性を検証した.

  34. Deep learning increases the availability of organism photographs taken by citizens in citizen science programs

    Yukari Suzuki-Ohno, Thomas Westfechtel, Jun Yokoyama, Kazunori Ohno, Tohru Nakashizuka, Masakado Kawata, Takayuki Okatani

    SCIENTIFIC REPORTS 12 (1) 2022年1月

    DOI: 10.1038/s41598-022-05163-5  

    ISSN:2045-2322

  35. Single-image Defocus Deblurring by Integration of Defocus Map Prediction Tracing the Inverse Problem Computation.

    Qian Ye, Masanori Suganuma, Takayuki Okatani

    CoRR abs/2207.03047 2022年

    DOI: 10.48550/arXiv.2207.03047  

  36. Learning Regularized Multi-Scale Feature Flow for High Dynamic Range Imaging.

    Qian Ye, Masanori Suganuma, Jun Xiao, Takayuki Okatani

    CoRR abs/2207.02539 2022年

    DOI: 10.48550/arXiv.2207.02539  

  37. Rethinking Unsupervised Domain Adaptation for Semantic Segmentation.

    Zhijie Wang, Masanori Suganuma, Takayuki Okatani

    CoRR abs/2207.00067 2022年

    DOI: 10.48550/arXiv.2207.00067  

  38. Learning to Bundle-adjust: A Graph Network Approach to Faster Optimization of Bundle Adjustment for Vehicular SLAM

    Tetsuya Tanaka, Yukihiro Sasagawa, Takayuki Okatani

    2021 IEEE/CVF International Conference on Computer Vision (ICCV) 6230-6239 2021年10月

    出版者・発行元: IEEE

    DOI: 10.1109/iccv48922.2021.00619  

  39. Acceleration-Based Deep Learning Method for Vehicle Monitoring

    Yanjie Zhu, Hidehiko Sekiya, Takayuki Okatani, Ikumasa Yoshida, Shuichi Hirano

    IEEE SENSORS JOURNAL 21 (15) 17154-17161 2021年8月

    DOI: 10.1109/JSEN.2021.3082145  

    ISSN:1530-437X

    eISSN:1558-1748

  40. 耐候性鋼板のさび外観評点識別精度に及ぼすCNNモデルと画像サイズの影響

    田井 政行, 関屋 英彦, 岡谷 貴之, 中村 聖三, 清水 隆史

    AI・データサイエンス論文集 2 (J2) 378-385 2021年

    出版者・発行元: 公益社団法人 土木学会

    DOI: 10.11532/jsceiii.2.j2_378  

    eISSN:2435-9262

    詳細を見る 詳細を閉じる

    耐候性鋼橋梁の点検・診断は,防食機能の劣化状態の判定に基づき行われており,その判定には外観評点法が用いられる.しかしながら,外観評点法により的確な評価を行うためには,相応の経験を要するが,昨今の人材不足や点検費用の確保などの課題があり,簡易かつ精度が高い評価手法の確立が求められている.本研究では,耐候性鋼橋梁のさび近接画像と既存の CNNモデルを活用し,外観評点の識別精度について検討を行った.また,識別精度に及ぼす学習・検証用近接画像の画像サイズの影響についても検討を行った.その結果,VGG19及び SEnetの CNNモデルが高い識別精度を示した.また,入力画像サイズが大きいほど識別精度が向上することを明らかにした.さらに,学習と検証に用いた画像の解像度が異なる場合,識別精度が低下する傾向があることを示した.

  41. Progressive and Selective Fusion Network for High Dynamic Range Imaging.

    Qian Ye, Jun Xiao, Kin-Man Lam 0001, Takayuki Okatani

    MM '21: ACM Multimedia Conference 5290-5297 2021年

    出版者・発行元: ACM

    DOI: 10.1145/3474085.3475651  

  42. Look Wide and Interpret Twice: Improving Performance on Interactive Instruction-following Tasks.

    Van-Quang Nguyen, Masanori Suganuma, Takayuki Okatani

    Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence(IJCAI) 923-930 2021年

    出版者・発行元: ijcai.org

    DOI: 10.24963/ijcai.2021/128  

  43. Matching in the Dark: A Dataset for Matching Image Pairs of Low-light Scenes.

    Wenzheng Song, Masanori Suganuma, Xing Liu 0010, Noriyuki Shimobayashi, Daisuke Maruta, Takayuki Okatani

    2021 IEEE/CVF International Conference on Computer Vision(ICCV) 6009-6018 2021年

    出版者・発行元: IEEE

    DOI: 10.1109/ICCV48922.2021.00597  

  44. Symbolizing Visual Features for Pre-training with Unlabeled Images.

    Yuichi Kamata, Moyuru Yamada, Keizo Kato, Akira Nakagawa, Takayuki Okatani

    Pattern Recognition - 6th Asian Conference 490-503 2021年

    出版者・発行元: Springer

    DOI: 10.1007/978-3-031-02444-3_37  

  45. Innovative technologies for infrastructure construction and maintenance through collaborative robots based on an open design approach.

    Keiji Nagatani, Masato Abe, Koichi Osuka, Pang-jo Chun, Takayuki Okatani, Mayuko Nishio, Shota Chikushi, Takamitsu Matsubara, Yusuke Ikemoto, Hajime Asama

    Advanced Robotics 35 (11) 715-722 2021年

    DOI: 10.1080/01691864.2021.1929471  

  46. Deep Learning for Channel-Agnostic Brain Decoding across Multiple Subjects

    Hiroto Date, Keisuke Kawasaki, Isao Hasegawa, Takayuki Okatani

    8th International Winter Conference on Brain-Computer Interface, BCI 2020 2020年2月1日

    出版者・発行元: Institute of Electrical and Electronics Engineers Inc.

    DOI: 10.1109/BCI48061.2020.9061615  

  47. Toward Explainable Fashion Recommendation. 査読有り

    Pongsate Tangseng, Takayuki Okatani

    IEEE Winter Conference on Applications of Computer Vision(WACV) 2142-2151 2020年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV45572.2020.9093367  

  48. Analysis and a Solution of Momentarily Missed Detection for Anchor-based Object Detectors.

    Yusuke Hosoya, Masanori Suganuma, Takayuki Okatani

    IEEE Winter Conference on Applications of Computer Vision(WACV) 1399-1407 2020年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV45572.2020.9093553  

  49. Efficient Attention Mechanism for Visual Dialog that Can Handle All the Interactions Between Multiple Inputs.

    Van-Quang Nguyen, Masanori Suganuma, Takayuki Okatani

    Computer Vision - ECCV 2020 - 16th European Conference 223-240 2020年

    出版者・発行元: Springer

    DOI: 10.1007/978-3-030-58586-0_14  

  50. Removal of Image Obstacles for Vehicle-mounted Surrounding Monitoring Cameras by Real-time Video Inpainting.

    Yoshihiro Hirohashi, Kenichi Narioka, Masanori Suganuma, Xing Liu 0010, Yukimasa Tamatsu, Takayuki Okatani

    2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition 857-866 2020年

    出版者・発行元: Computer Vision Foundation / IEEE

    DOI: 10.1109/CVPRW50498.2020.00115  

  51. Hyperparameter-Free Out-of-Distribution Detection Using Cosine Similarity.

    Engkarat Techapanurak, Masanori Suganuma, Takayuki Okatani

    Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision 53-69 2020年

    出版者・発行元: Springer

    DOI: 10.1007/978-3-030-69538-5_4  

  52. Extending information maximization from a rate-distortion perspective.

    Yan Zhang 0055, Junjie Hu, Takayuki Okatani

    Neurocomputing 399 285-295 2020年

    DOI: 10.1016/j.neucom.2020.02.061  

    ISSN:0925-2312

    eISSN:1872-8286

  53. A vision-based method for crack detection in gusset plate welded joints of steel bridges using deep convolutional neural networks

    Cao Vu Dung, Hidehiko Sekiya, Suichi Hirano, Takayuki Okatani, Chitoshi Miki

    Automation in Construction 102 217-229 2019年6月1日

    DOI: 10.1016/j.autcon.2019.02.013  

    ISSN:0926-5805

  54. Attention-based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions. 査読有り

    Masanori Suganuma, Xing Liu, Takayuki Okatani

    Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019年6月

  55. Multi-task Learning of Hierarchical Vision-Language Representation. 査読有り

    Duy-Kien Nguyen, Takayuki Okatani

    Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019年6月

  56. Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration. 査読有り

    Xing Liu, Masanori Suganuma, Zhun Sun, Takayuki Okatani

    Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019年6月

  57. A Generative Model of Underwater Images for Active Landmark Detection and Docking.

    Shuang Liu 0002, Mete Özay, Hongli Xu, Yang Lin, Takayuki Okatani

    2019 IEEE/RSJ International Conference on Intelligent Robots and Systems(IROS) 8034-8039 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/IROS40897.2019.8968146  

  58. Visualization of Convolutional Neural Networks for Monocular Depth Estimation.

    Junjie Hu, Yan Zhang 0055, Takayuki Okatani

    2019 IEEE/CVF International Conference on Computer Vision(ICCV) 3868-3877 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/ICCV.2019.00397  

  59. Improving Head Pose Estimation with a Combined Loss and Bounding Box Margin Adjustment.

    Mingzhen Shao, Zhun Sun, Mete Ozay, Takayuki Okatani

    14th IEEE International Conference on Automatic Face & Gesture Recognition(FG) 1-5 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/FG.2019.8756605  

  60. Deep Learning for Natural Image Reconstruction from Electrocorticography Signals.

    Hiroto Date, Keisuke Kawasaki, Isao Hasegawa, Takayuki Okatani

    2019 IEEE International Conference on Bioinformatics and Biomedicine(BIBM) 2331-2336 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/BIBM47256.2019.8983029  

  61. ImPACT-TRC Thin Serpentine Robot Platform for Urban Search and Rescue. 査読有り

    Masashi Konyo, Yuichi Ambe, Hikaru Nagano, Yu Yamauchi, Satoshi Tadokoro, Yoshiaki Bando, Katsutoshi Itoyama, Hiroshi G. Okuno, Takayuki Okatani, Kanta Shimizu, Eisuke Ito

    Disaster Robotics - Results from the ImPACT Tough Robotics Challenge 25-76 2019年

    出版者・発行元: Springer

    DOI: 10.1007/978-3-030-05321-5_2  

  62. Video-Rate Video Inpainting. 査読有り

    Rito Murase, Yan Zhang, Takayuki Okatani

    IEEE Winter Conference on Applications of Computer Vision, WACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019 1553-1561 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV.2019.00170  

  63. Revisiting Single Image Depth Estimation: Toward Higher Resolution Maps With Accurate Object Boundaries. 査読有り

    Junjie Hu, Mete Ozay, Yan Zhang, Takayuki Okatani

    IEEE Winter Conference on Applications of Computer Vision, WACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019 1043-1051 2019年

    出版者・発行元: IEEE

    DOI: 10.1109/WACV.2019.00116  

  64. Improving Generalization Ability of Deep Neural Networks for Visual Recognition Tasks. 査読有り

    Takayuki Okatani, Xing Liu 0010, Masanori Suganuma

    Computational Color Imaging - 7th International Workshop, CCIW 2019, Chiba, Japan, March 27-29, 2019, Proceedings 3-13 2019年

    出版者・発行元: Springer

    DOI: 10.1007/978-3-030-13940-7_1  

  65. Detection and Pose Estimation for Short-Range Vision-Based Underwater Docking. 査読有り

    Shuang Liu 0002, Mete Ozay, Takayuki Okatani, Hongli Xu, Kai Sun, Yang Lin

    IEEE Access 7 2720-2749 2019年

    DOI: 10.1109/ACCESS.2018.2885537  

  66. Learning deep representations and detection of docking stations using underwater imaging

    Shuang Liu, Mete Ozay, Takayuki Okatani, Hongli Xu, Yang Lin, Haitao Gu

    2018 OCEANS - MTS/IEEE Kobe Techno-Oceans, OCEANS - Kobe 2018 2018年12月4日

    出版者・発行元: Institute of Electrical and Electronics Engineers Inc.

    DOI: 10.1109/OCEANSKOBE.2018.8559067  

  67. Recommending Outfits from Personal Closet 査読有り

    Pongsate Tangseng, Kota Yamaguchi, Takayuki Okatani

    Proceedings - 2017 IEEE International Conference on Computer Vision Workshops, ICCVW 2017 2018- 2275-2279 2018年1月19日

    出版者・発行元: Institute of Electrical and Electronics Engineers Inc.

    DOI: 10.1109/ICCVW.2017.267  

  68. Perturbation Tolerance of Deep Neural Networks and Humans in Material Recognition 査読有り

    Xing Liu, Masataka Sawayama, Ryusuke Hayashi, Mete Ozay, Takayuki Okatani, Shin'ya Nishida

    Journal of Vision 18 (10) 756-756 2018年

    DOI: 10.1167/18.10.756  

  69. Deep Structured Energy-Based Image Inpainting. 査読有り

    Fazil Altinel, Mete Ozay, Takayuki Okatani

    24th International Conference on Pattern Recognition, ICPR 2018, Beijing, China, August 20-24, 2018 423-428 2018年

    出版者・発行元: IEEE Computer Society

    DOI: 10.1109/ICPR.2018.8546025  

  70. Exploiting the Potential of Standard Convolutional Autoencoders for Image Restoration by Evolutionary Search. 査読有り

    Masanori Suganuma, Mete Ozay, Takayuki Okatani

    Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15, 2018 4778-4787 2018年

    出版者・発行元: JMLR.org

  71. Feature Quantization for Defending Against Distortion of Images. 査読有り

    Zhun Sun, Mete Ozay, Yan Zhang, Xing Liu, Takayuki Okatani

    2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 7957-7966 2018年

    出版者・発行元: IEEE Computer Society

  72. Improved Fusion of Visual and Language Representations by Dense Symmetric Co-Attention for Visual Question Answering. 査読有り

    Duy-Kien Nguyen, Takayuki Okatani

    2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 6087-6096 2018年

    出版者・発行元: IEEE Computer Society

  73. Training CNNs With Normalized Kernels. 査読有り

    Mete Özay, Takayuki Okatani

    Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orl 3884-3891 2018年

    出版者・発行元: AAAI Press

  74. Truncating Wide Networks Using Binary Tree Architectures 査読有り

    Yan Zhangy, Mete Ozayy, Shuohao Li, Takayuki Okatani

    Proceedings of the IEEE International Conference on Computer Vision 2017- 2116-2124 2017年12月22日

    DOI: 10.1109/ICCV.2017.231  

    ISSN:1550-5499

  75. Attention to describe products with attributes 査読有り

    Shuohao Li, Kota Yamaguchi, Takayuki Okatani

    Proceedings of the 15th IAPR International Conference on Machine Vision Applications, MVA 2017 215-218 2017年7月19日

    出版者・発行元: Institute of Electrical and Electronics Engineers Inc.

    DOI: 10.23919/MVA.2017.7986839  

  76. Temporal city modeling using street level imagery 査読有り

    Ken Sakurada, Daiki Tetsuka, Takayuki Okatani

    COMPUTER VISION AND IMAGE UNDERSTANDING 157 55-71 2017年4月

    DOI: 10.1016/j.cviu.2017.01.012  

    ISSN:1077-3142

    eISSN:1090-235X

  77. Learning to describe E-commerce images from noisy online data 査読有り

    Takuya Yashima, Naoaki Okazaki, Kentaro Inui, Kota Yamaguchi, Takayuki Okatani

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 10115 85-100 2017年

    DOI: 10.1007/978-3-319-54193-8_6  

    ISSN:1611-3349 0302-9743

    eISSN:1611-3349

  78. Self-calibration-based Approach to Critical Motion Sequences of Rolling-shutter Structure from Motion 査読有り

    Eisuke Ito, Takayuki Okatani

    30TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017) 4512-4520 2017年

    DOI: 10.1109/CVPR.2017.480  

    ISSN:1063-6919

  79. Self-calibration-based Approach to Critical Motion Sequences of Rolling-shutter Structure from Motion 査読有り

    Eisuke Ito, Takayuki Okatani

    30TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017) 4512-4520 2017年

    DOI: 10.1109/CVPR.2017.480  

    ISSN:1063-6919

  80. Separation of reflection components by sparse non-negative matrix factorization 査読有り

    Yasushi Akashi, Takayuki Okatani

    COMPUTER VISION AND IMAGE UNDERSTANDING 146 77-85 2016年5月

    DOI: 10.1016/j.cviu.2015.09.001  

    ISSN:1077-3142

    eISSN:1090-235X

  81. Hybrid macro-micro visual analysis for city-scale state estimation 査読有り

    Ken Sakurada, Takayuki Okatani, Kris M. Kitani

    COMPUTER VISION AND IMAGE UNDERSTANDING 146 86-98 2016年5月

    DOI: 10.1016/j.cviu.2016.02.017  

    ISSN:1077-3142

    eISSN:1090-235X

  82. A Gaze-Reactive Display for Simulating Depth-of-Field of Eyes When Viewing Scenes with Multiple Depths 査読有り

    Tatsuro Orikasa, Takayuki Okatani

    IEICE TRANSACTIONS ON INFORMATION AND SYSTEMS E99D (3) 739-746 2016年3月

    DOI: 10.1587/transinf.2015EDP7110  

    ISSN:1745-1361

  83. Design of Kernels in Convolutional Neural Networks for Image Classification 査読有り

    Zhun Sun, Mete Ozay, Takayuki Okatani

    COMPUTER VISION - ECCV 2016, PT VII 9911 51-66 2016年

    DOI: 10.1007/978-3-319-46478-7_4  

    ISSN:0302-9743

  84. Automatic Attribute Discovery with Neural Activations 査読有り

    Sirion Vittayakorn, Takayuki Umeda, Kazuhiko Murasaki, Kyoko Sudo, Takayuki Okatani, Kota Yamaguchi

    COMPUTER VISION - ECCV 2016, PT IV 9908 252-268 2016年

    DOI: 10.1007/978-3-319-46493-0_16  

    ISSN:0302-9743

  85. Recognizing open-vocabulary relations between objects in images 査読有り

    Muraoka, M, Maharjan, S, Saito, M, Yamaguchi, K, Okazaki, N, Okatani, T, Inui, K

    Proceedings of the 30th Pacific Asia Conference on Language, Information and Computation, PACLIC 2016 239-248 2016年

  86. Integrating Deep Features for Material Recognition 査読有り

    Yan Zhang, Mete Ozay, Xing Liu, Takayuki Okatani

    2016 23RD INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR) 3697-3702 2016年

    DOI: 10.1109/ICPR.2016.7900209  

    ISSN:1051-4651

  87. Possibility to use product image and review text based on the association between onomatopoeia and texture 査読有り

    Ryuichi Doizaki, Saki Iiba, Takayuki Okatani, Maki Sakamoto

    Transactions of the Japanese Society for Artificial Intelligence 30 (1) 124-137 2015年1月6日

    出版者・発行元: Japanese Society for Artificial Intelligence

    DOI: 10.1527/tjsai.30.124  

    ISSN:1346-8030 1346-0714

  88. Change Detection from a Street Image Pair using CNN Features and Superpixel Segmentation.

    Ken Sakurada, Takayuki Okatani

    Proceedings of the British Machine Vision Conference 2015(BMVC) 61-12 2015年

    出版者・発行元: BMVA Press

    DOI: 10.5244/C.29.61  

  89. Transformation of Markov Random Fields for Marginal Distribution Estimation 査読有り

    Masaki Saito, Takayuki Okatani

    2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) 07-12-June-2015 797-805 2015年

    DOI: 10.1109/CVPR.2015.7298680  

    ISSN:1063-6919

  90. Detecting Building-level Changes of a City Using Street Images and a 2D City Map 査読有り

    Daiki Tetsuka, Takayuki Okatani

    2015 IEEE Winter Conference on Applications of Computer Vision (WACV) 349-356 2015年

    DOI: 10.1109/WACV.2015.53  

  91. Separation of Reflection Components by Sparse Non-negative Matrix Factorization 査読有り

    Yasuhiro Akashi, Takayuki Okatani

    COMPUTER VISION - ACCV 2014, PT V 9007 611-625 2015年

    DOI: 10.1007/978-3-319-16814-2_40  

    ISSN:0302-9743

  92. Understanding Convolutional Neural Networks in Terms of Category-Level Attributes 査読有り

    Makoto Ozeki, Takayuki Okatani

    COMPUTER VISION - ACCV 2014, PT II 9004 362-375 2015年

    DOI: 10.1007/978-3-319-16808-1_25  

    ISSN:0302-9743

  93. Massive City-Scale Surface Condition Analysis Using Ground and Aerial Imagery 査読有り

    Ken Sakurada, Takayuki Okatani, Kris M. Kitani

    COMPUTER VISION - ACCV 2014, PT I 9003 49-64 2015年

    DOI: 10.1007/978-3-319-16865-4_4  

    ISSN:0302-9743

  94. Creating multi-viewpoint panoramas of streets with sparsely located buildings 査読有り

    Takayuki Okatani, Jun Yanagisawa, Daiki Tetsuka, Ken Sakurada, Koichiro Deguchi

    Springer Tracts in Advanced Robotics 92 65-79 2014年

    DOI: 10.1007/978-3-642-40686-7_5  

    ISSN:1610-7438 1610-742X

  95. Improving accuracy of planar tracking by resolving resolution inconsistencies 査読有り

    Tomohiro Ushiki, Eisuke Ito, Takayuki Okatani

    IPSJ Transactions on Computer Vision and Applications 5 109-113 2013年7月

    DOI: 10.2197/ipsjtcva.5.109  

    ISSN:1882-6695

  96. Detecting Changes in 3D Structure of a Scene from Multi-view Images Captured by a Vehicle-mounted Camera 査読有り

    Ken Sakurada, Takayuki Okatani, Koichiro Deguchi

    2013 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) 137-144 2013年

    DOI: 10.1109/CVPR.2013.25  

    ISSN:1063-6919

  97. Discrete MRF Inference of Marginal Densities for Non-uniformly Discretized Variable Space 査読有り

    Masaki Saito, Takayuki Okatani, Koichiro Deguchi

    2013 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) 57-64 2013年

    DOI: 10.1109/CVPR.2013.15  

    ISSN:1063-6919

  98. Application of the Mean Field Methods to MRF Optimization in Computer Vision 査読有り

    Masaki Saito, Takayuki Okatani, Koichiro Deguchi

    2012 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) 1680-1687 2012年

    DOI: 10.1109/CVPR.2012.6247862  

    ISSN:1063-6919

  99. Optimal Integration of Photometric and Geometric Surface Measurements Using Inaccurate Reflectance/Illumination Knowledge 査読有り

    Takayuki Okatani, Koichiro Deguchi

    2012 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) 254-261 2012年

    DOI: 10.1109/CVPR.2012.6247683  

    ISSN:1063-6919

  100. Recognizing Surface Qualities from Natural Images based on Learning to Rank 査読有り

    Takashi Abe, Takayuki Okatani, Koichiro Deguchi

    2012 21ST INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR 2012) 3712-3715 2012年

    ISSN:1051-4651

  101. Optimum method for real-time reconstruction of sensor surface in total-internal-reflection based tactile sensor. 査読有り

    Ryosuke Taira, Satoshi Saga, Takayuki Okatani, Koichiro Deguchi

    IEEE World Haptics Conference, WHC 2011, 21-24 June 2011, Istanbul, Turkey 329-334 2011年

    出版者・発行元: IEEE Computer Society

    DOI: 10.1109/WHC.2011.5945507  

  102. Flexible Online Calibration for a Mobile Projector-Camera System 査読有り

    Daisuke Abe, Takayuki Okatani, Koichiro Deguchi

    COMPUTER VISION - ACCV 2010, PT IV 6495 565-579 2011年

    DOI: 10.1007/978-3-642-19282-1_45  

    ISSN:0302-9743

  103. Hand Gesture Recognition using Histogram of Oriented Gradients and Partial Least Squares Regression 査読有り

    Arindam Misra, Abe Takashi, Takayuki Okatani, Koichiro Deguchi

    Proceedings of IAPR Conference on Machine Vision Applications 479-482 2011年

  104. Sensing method of total-internal-reflection-based tactile sensor 査読有り

    Momotaro Koike, Satoshi Saga, Takayuki Okatani, Koichiro Deguchi

    2011 IEEE World Haptics Conference, WHC 2011 615-619 2011年

    DOI: 10.1109/WHC.2011.5945556  

  105. Accurate and Robust Planar Tracking based on a Model of Image Sampling and Reconstruction Process 査読有り

    Eisuke Ito, Takayuki Okatani, Koichiro Deguchi

    2011 10TH IEEE INTERNATIONAL SYMPOSIUM ON MIXED AND AUGMENTED REALITY (ISMAR) 2011年

    DOI: 10.1109/ISMAR.2011.6092364  

    ISSN:1554-7868

  106. Efficient Algorithm for Low-rank Matrix Factorization with Missing Components and Performance Comparison of Latest Algorithms 査読有り

    Takayuki Okatani, Takahiro Yoshida, Koichiro Deguchi

    2011 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV) 842-849 2011年

    DOI: 10.1109/ICCV.2011.6126324  

    ISSN:1550-5499

  107. 3D reconstruction of reflective surface on reflection type tactile sensor using constraints of geometrical optics

    Ryosuke Taira, Satoshi Saga, Takayuki Okatani, Koichiro Deguchi

    Proceedings of the SICE Annual Conference 3144-3149 2010年

    出版者・発行元: Society of Instrument and Control Engineers (SICE)

  108. A projector-camera system for high-quality synthesis of virtual reflectance on real object surfaces 査読有り

    Tomoya Okazaki, Takayuki Okatani, Koichiro Deguchi

    IPSJ Transactions on Computer Vision and Applications 2 71-83 2010年

    DOI: 10.2197/ipsjtcva.2.71  

    ISSN:1882-6695

  109. Easy calibration of a multi-projector display system 査読有り

    Takayuki Okatani, Koichiro Deguchi

    International Journal of Computer Vision 85 (1) 1-18 2009年4月

    DOI: 10.1007/s11263-009-0242-0  

  110. 被写界深度ぼけの提示により奥行き感を強化する注視反応型ディスプレイ 査読有り

    岡谷貴之, 石澤昴, 出口光一郎

    電子情報通信学会論文誌 J92-D (8) 1298-1307 2009年3月

  111. Study of Image Quality of Superimposed Projection Using Multiple Projectors 査読有り

    Takayuki Okatani, Mikio Wada, Koichiro Deguchi

    IEEE TRANSACTIONS ON IMAGE PROCESSING 18 (2) 424-429 2009年2月

    DOI: 10.1109/TIP.2008.2008899  

    ISSN:1057-7149

  112. Modeling effects of urban spatial structure on urban thermal environment - Integrating satellite image processing and metric model analysis

    Nguyen Thi Mai Dung, Koichiro Deguchi, Takayuki Okatani

    30th Asian Conference on Remote Sensing 2009, ACRS 2009 1 590-596 2009年

  113. Shape Reconstruction by Combination of Structured-Light Projection and Photometric Stereo Using a Projector-Camera System High Quality Reproduction of a Virtual Reflectance Property on a Real Object Surface 査読有り

    Tomoya Okazaki, Takayuki Okatani, Koichiro Deguchi

    ADVANCES IN IMAGE AND VIDEO TECHNOLOGY, PROCEEDINGS 5414 410-422 2009年

    ISSN:0302-9743

  114. On Bias Correction for Geometric Parameter Estimation in Computer Vision 査読有り

    Takayuki Okatani, Koichiro Deguchi

    CVPR: 2009 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOLS 1-4 959-966 2009年

    DOI: 10.1109/CVPRW.2009.5206722  

    ISSN:1063-6919

  115. Improving Accuracy of Geometric Parameter Estimation Using Projected Score Method 査読有り

    Takayuki Okatani, Koichiro Deguchi

    2009 IEEE 12TH INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV) 1733-1740 2009年

    DOI: 10.1109/ICCV.2009.5459388  

    ISSN:1550-5499

  116. Absolute phase measurements using geometric constraints between multiple cameras and projectors 査読有り

    Rui Ishiyama, Shizuo Sakamoto, Johji Tajima, Takayuki Okatani, Koichiro Deguchi

    Applied Optics 46 (17) 3528-3539 2007年6月10日

    出版者・発行元: OSA - The Optical Society

    DOI: 10.1364/AO.46.003528  

    ISSN:1539-4522 1559-128X

    eISSN:1539-4522

  117. On the Wiberg algorithm for matrix factorization in the presence of missing components 査読有り

    Takayuki Okatani, Koichiro Deguchi

    INTERNATIONAL JOURNAL OF COMPUTER VISION 72 (3) 329-337 2007年5月

    DOI: 10.1007/s11263-006-9785-5  

    ISSN:0920-5691

  118. 3D spatial information extraction from ALOS PRISM triple image

    Nguyen Thi Mai Dung, Koichiro Deguchi, Takayuki Okatani

    28th Asian Conference on Remote Sensing 2007, ACRS 2007 2 1199-1203 2007年

  119. A Physics-Based Imaging Model of Scanning Electron Microscopes 査読有り

    Kousuke Kamada, Takayuki Okatani, Koichiro Deguchi

    Proceedings of IAPR Conference on Machine Vision Applications 102-105 2007年

  120. Variational Bayes based approach to robust subspace learning 査読有り

    Takayuki Okatani, Koichiro Deguchi

    2007 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOLS 1-8 1004-+ 2007年

    DOI: 10.1109/CVPR.2007.383101  

    ISSN:1063-6919

  121. Estimating scale of a scene from a single image based on defocus blur and scene geometry 査読有り

    Takayuki Okatani, Koichiro Deguchi

    2007 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOLS 1-8 1890-+ 2007年

    DOI: 10.1109/CVPR.2007.383217  

    ISSN:1063-6919

  122. Precise 3-D measurement using uncalibrated pattern projection 査読有り

    Rui Ishiyama, Takayuki Okatani, Koichiro Deguchi

    2007 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING, VOLS 1-7 (1) 225-228 2007年

    DOI: 10.1109/ICIP.2007.4378932  

    ISSN:1522-4880

  123. HHMM based recognition of human activity 査読有り

    Daiki Kawanaka, Takayuki Okatani, Koichiro Deguchi

    IEICE TRANSACTIONS ON INFORMATION AND SYSTEMS E89D (7) 2180-2185 2006年7月

    DOI: 10.1093/ietisy/e89-d.7.2180  

    ISSN:1745-1361

  124. Video synchronization based on co-occurrence of appearance changes in video sequences 査読有り

    Manabu Ushizaki, Takayuki Okatani, Kochiro Deguchi

    18TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3, PROCEEDINGS 71-+ 2006年

    DOI: 10.1109/ICPR.2006.1174  

    ISSN:1051-4651

  125. The importance of gaze control mechanism on vision-based motion control of a biped robot 査読有り

    Shun Ushida, Kousuke Yoshimi, Takayuki Okatani, Koichiro Deguchi

    2006 IEEE/RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS, VOLS 1-12 4447-+ 2006年

    DOI: 10.1109/IROS.2006.282079  

  126. Autocalibration of an ad hoc construction of multi-projector displays 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2006 2006年

    DOI: 10.1109/CVPRW.2006.35  

    ISSN:1063-6919

  127. Autocalibration of a projector-camera system 査読有り

    T Okatani, K Deguchi

    IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 27 (12) 1845-1855 2005年12月

    DOI: 10.1109/TPAMI.2005.235  

    ISSN:0162-8828

    eISSN:1939-3539

  128. 行動と画像変化を結びつけたロボットの空間認識 査読有り

    中川智洋, 岡谷貴之, 出口光一郎

    日本ロボット学会誌 23 (3) 90-97 2005年3月

    出版者・発行元: The Robotics Society of Japan

    DOI: 10.7210/jrsj.23.362  

    ISSN:0289-1824

    詳細を見る 詳細を閉じる

    In this paper, we propose an active vision strategy for the construction of a 3D map in a robot brain from its stereo eye images. We show that the combination of the robot action and image change caused by the action improves the accuracy of the map. If the robot stereo cameras have been accurately calibrated, the obtained reconstruction of the static scene stays stationary. If the calibration is not accurate, the reconstructed scene changes according to the robot action. Therefore, we utilize this change to correct the parameters of the stereo cameras and the 3D map so as to obtain stationary scene under the action. We show the feasibility of this idea as an auto-calibration of robot vision with some simulation experiments and implementation on actual robots.

  129. HHMM based recognition of human activity from motion trajectories in image sequences

    Daiki Kawanaka, Shun Ushida, Takayuki Okatani, Koichiro Deguchi

    Proceedings of the 9th IAPR Conference on Machine Vision Applications, MVA 2005 578-581 2005年

  130. Markerless augmented reality using planar structures

    Hajime Tomizawa, Takayuki Okatani, Koichiro Deguchi

    Proceedings of the SICE Annual Conference 1066-1070 2005年

  131. A method for 3D shape reconstruction from SEM images

    Kosuke Kamada, Koichiro Deguchi, Takayuki Okatani

    Proceedings of the SICE Annual Conference 650-653 2005年

  132. Reconstruction of the spatial motion and posture of tools from stereo image sequence

    Manabu Ushizaki, Takayuki Okatani, Koichiro Deguchi

    Proceedings of the SICE Annual Conference 644-648 2005年

  133. Motion Stabilization of Biped Robot by Gaze Control 査読有り

    Shota Takizawa, Shun Ushida, Takayuki Okatani, Koichiro Deguchi

    Proceedings of IAPR Conference on Machine Vision Applications 324-327 2005年

  134. 2DOF motion stabilization of biped robot by gaze control strategy 査読有り

    S Takizawa, S Ushida, T Okatani, K Deguchi

    2005 IEEE/RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS, VOLS 1-4 3809-3814 2005年

    DOI: 10.1109/IROS.2005.1545562  

  135. Contrast enhancement of fundus images using regional histograms for medical diagnosis

    Tatsuya Shimahara, Takayuki Okatani, Koichiro Deguchi

    Proceedings of the SICE Annual Conference 411-414 2004年

  136. High-speed and high-precision optical flow detection for real-time motion segmentation 査読有り

    H Ishiyama, T Okatani, K Deguchi

    SICE 2004 ANNUAL CONFERENCE, VOLS 1-3 1202-1205 2004年

  137. A probabilistic approach to linear subspace fitting for computer vision problems 査読有り

    Takayuki Okatani

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops 2004- (January) 2004年

    出版者・発行元: IEEE Computer Society

    DOI: 10.1109/CVPR.2004.286  

    ISSN:2160-7516 2160-7508

  138. Active construction of 3D map in robot by combining motion and perceived images 査読有り

    Tomohiro Nakagawa, Takayuki Okatani, Koichiro Deguchi

    Proceedings of Asian Conference on Computer Vision 1 563-568 2004年

  139. Object tracking by the mean-shift of regional color distribution combined with the particle-filter algorithm 査読有り

    K Deguchi, O Kawanaka, T Okatani

    PROCEEDINGS OF THE 17TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3 506-509 2004年

    DOI: 10.1109/ICPR.2004.1334577  

    ISSN:1051-4651

  140. A Color-based Probabilistic Tracking by Using Graphical Models 査読有り

    Y. Sato, T. Okatani, K. Deguchi

    Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems 631-636 2004年

  141. A color-based tracking by Kalman particle filter 査読有り

    Y Satoh, T Okatani, K Deguchi

    PROCEEDINGS OF THE 17TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3 502-505 2004年

    DOI: 10.1109/ICPR.2004.1334576  

    ISSN:1051-4651

  142. 実環境での物体認識のための距離画像マッチング 査読有り

    長尾影洋, 岡谷貴之, 出口光一郎

    計測自動制御学会論文誌 40 (1) 1-8 2004年1月

    出版者・発行元: The Society of Instrument and Control Engineers

    DOI: 10.9746/sicetr1965.40.10  

    ISSN:0453-4654

    詳細を見る 詳細を閉じる

    Object recognition is very important task in the field of computer vision. In this paper, we present a range image matching for the object recognition in real scene. The matching is the process of comparing a range image with a stored shape data. We first discuss that the difference of resolutions of them must be considered for the range image matching. Then, we analyze the apparent object shape change of the range image due to the measurement distance in the relation to the property of the range finder. We claim that the apparent shape change can be modelled by a smoothing process of the object surface. We propose a method of matching after shape smoothing of the stored data to cope with has a very low resolution of the range image. We also propose a generation of a surface model of the given data, considering the resolution difference of a newly obtained data of the target object. Finally, we show several experimental results which promises the feasibility of the proposal method.

  143. Comparison of the feature point tracking method in image sequences 査読有り

    Y Naito, T Okatani, K Deguchi

    SICE 2003 ANNUAL CONFERENCE, VOLS 1-3 1326-1331 2003年

  144. An extraction method of moving objects by shape energy functions with prior knowledge 査読有り

    Y Satoh, T Okatani, K Deguchi

    SICE 2003 ANNUAL CONFERENCE, VOLS 1-3 1320-1325 2003年

  145. Tracking multiple three-dimensional motions by using modified condensation algorithm and multiple images 査読有り

    K Hamasaki, T Nakajima, T Okatani, K Deguchi

    IROS 2003: PROCEEDINGS OF THE 2003 IEEE/RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS, VOLS 1-4 236-241 2003年

  146. Autocalibration of a projector-sereen-camera system: Theory and algorithm for screen-to-camera homography estimation 査読有り

    T Okatani, K Deguchi

    NINTH IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION, VOLS I AND II, PROCEEDINGS 774-781 2003年

  147. Binocular motion tracking by gaze fixation control and three-dimensional shape reconstruction 査読有り

    Y Satoh, T Okatani, K Deguchi

    ADVANCED ROBOTICS 17 (10) 1057-1072 2003年

    DOI: 10.1163/156855303322554427  

    ISSN:0169-1864

    eISSN:1568-5535

  148. Toward a statistically optimal method for estimating geometric relations from noisy data: Cases of linear relations 査読有り

    T Okatani, K Deguchi

    2003 IEEE COMPUTER SOCIETY CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOL 1, PROCEEDINGS 432-439 2003年

    ISSN:1063-6919

  149. 3D object surface generation from range point data for object shape model matching 査読有り

    K Nagao, T Okatani, K Deguchi

    SICE 2002: PROCEEDINGS OF THE 41ST SICE ANNUAL CONFERENCE, VOLS 1-5 1416-1419 2002年

  150. Is There Room for Improving Estimation Accuracy of the Structure and Motion Problem? 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Workshop of Statistical Method in Video Processing 2002年

  151. Robust estimation of camera translation between two images using a camera with a 3D orientation sensor 招待有り 査読有り

    T Okatani, K Deguchi

    16TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL I, PROCEEDINGS 275-278 2002年

    ISSN:1051-4651

  152. 3-Dimensional Tracking of Multiple Object Motions from Multi-view Image by Mixed-State CONDENSATION Algorithm 査読有り

    Koichiro Deguchi, Koji Hamasaki, Taira Nakajima, Takayuki Okatani

    Proceedings of IAPR Workshop on Machine Vision Applications 408-411 2002年

  153. Theory of Nuisance Parameter Problems in Computer Vision: Several Examples of Application to Linear Cases 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Proceedings of the Workshop on Science of Vision 94-100 2002年

  154. A motion tracking by extracting 3D feature of moving objects with binocular cooperative fixation 査読有り

    Y Satoh, T Nakagawa, T Okatani, K Deguchi

    2002 IEEE/RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS, VOLS 1-3, PROCEEDINGS 13-18 2002年

  155. カメラモデルの段階的切り換えによる多視点カメラシステムのセルフキャリブレーション 査読有り

    岡谷貴之, 久保田俊作, 出口光一郎

    情報処理学会論文誌:コンピュータビジョンとイメージメディア 43 (11) 33-42 2002年

  156. On uniqueness of solutions of the three-light-source photometric stereo: Conditions on illumination configuration and surface reflectance 査読有り

    T Okatani, K Deguchi

    COMPUTER VISION AND IMAGE UNDERSTANDING 81 (2) 211-226 2001年2月

    DOI: 10.1006/cviu.2000.0887  

    ISSN:1077-3142

  157. On photometric aspects of catadioptric cameras 査読有り

    T Okatani, K Deguchi

    2001 IEEE COMPUTER SOCIETY CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, VOL 1, PROCEEDINGS 1106-1113 2001年

    ISSN:1063-6919

  158. 反射鏡を利用した全方位カメラの画像照度について 査読有り

    岡谷貴之, 出口光一郎

    情報処理学会論文誌:コンピュータビジョンとイメージメディア 42 (13) 71-80 2001年

  159. Closed form solution of local shape from shading at critical points 査読有り

    T Okatani, K Deguchi

    INTERNATIONAL JOURNAL OF COMPUTER VISION 40 (2) 169-178 2000年11月

    DOI: 10.1023/A:1026510403689  

    ISSN:0920-5691

    eISSN:1573-1405

  160. 非ランバート性拡散反射に対する等高線の発展による陰影からの形状復元 査読有り

    岡谷 貴之

    情報処理学会論文誌 41 (1) 92-100 2000年

    出版者・発行元: 一般社団法人情報処理学会

    ISSN:1882-7764

    詳細を見る 詳細を閉じる

    陰影から形状を復元する問題は1階の偏微分方程式として定式化される.解法には古典的な特性方程式による方法のほかに,等高線の発展によって解を計算する方法がある.従来のやり方ではランバート面が仮定されていた.本論文では,ランバート面に限らない一般の拡散反射を考慮する.等高線に基づく形状復元が可能になるために反射率分布図が満たすべき条件を示す.The problem of recovering shape from shading is formulated as solvingthe first order partial differential equation.Besides the classical method of the characteristic equation,a group of methods that computes the solution by trackingits equal hight contour has been proposed.But all of them assumed the Lambertian reflectance.This paper extends the methods to general diffuse reflectancethat is not necessarily Lambertian.We show a condition that the reflectance map should satisfyso that the methods can be used.

  161. Estimation of illumination distribution using a specular sphere

    T Okatani, K Deguchi

    15TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3, PROCEEDINGS 592-595 2000年

    ISSN:1051-4651

  162. Recovering Camera Motion from Image Sequence Based on Registration of Silhouette Cones : Shape from Silhouette Using a Mobile Camera with a Gyro Sensor 査読有り

    Takayuki Okatani, Koichiro Deguchi

    IAPR Workshop on Machine Vision Applications 451-454 2000年

  163. Estimation of illumination distribution using a specular sphere 査読有り

    T Okatani, K Deguchi

    15TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3, PROCEEDINGS 592-595 2000年

    ISSN:1051-4651

  164. Yet Another Appearance-Based Method for Pose Estimation Based on a Linear Model 査読有り

    Takayuki OKATANI, Koichiro DEGUCHI

    Proceedings of IAPR Conference on Machine Vision Applications 2000年

  165. Computation of the sign of the Gaussian curvature of a surface from multiple unknown illumination images without knowledge of the reflectance property 査読有り

    T Okatani, K Deguchi

    COMPUTER VISION AND IMAGE UNDERSTANDING 76 (2) 125-134 1999年11月

    DOI: 10.1006/cviu.1999.0792  

    ISSN:1077-3142

  166. 陰影画像における特異点に関する等濃度線の不変性について 査読有り

    岡谷 貴之

    情報処理学会論文誌 40 (9) 3446-3454 1999年

  167. 陰影と反射率分布図の整合性について

    岡谷貴之, 出口光一郎

    情報処理学会研究報告 117 (6) 39-46 1999年

  168. 複数の画像間の陰影差に基づく曲画のガウス曲率の符号の決定 査読有り

    岡谷 貴之

    情報処理学会論文誌 39 (6) 1965-1972 1998年

  169. On identification of singular points using photometric invariants for global shape from shading problem 査読有り

    T Okatani, K Deguchi

    FOURTEENTH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOLS 1 AND 2 1787-1790 1998年

    ISSN:1051-4651

  170. Shape reconstruction from an endoscope image by shape from shading technique for a point light source at the projection center 査読有り

    T Okatani, K Deguchi

    COMPUTER VISION AND IMAGE UNDERSTANDING 66 (2) 119-131 1997年5月

    DOI: 10.1006/cviu.1997.0613  

    ISSN:1077-3142

  171. 内視鏡画像からの陰影に基づく立体形状復元 査読有り

    岡谷 貴之

    計測自動制御学会論文集 33 (10) 1035-1042 1997年

    出版者・発行元: The Society of Instrument and Control Engineers

    DOI: 10.9746/sicetr1965.33.1035  

    ISSN:0453-4654

    詳細を見る 詳細を閉じる

    A new method for reconstructing three dimensional object from an endoscope image is presented. The proposed method uses image shading generated by a light source at the endoscope head. The problem of reconstructing three dimensional shape of an object from its shading has been widely studied in the literature of computer vision research. There have been many significant developments made in this area. However, most of the conventional methods are not applicable to this case, since they are based on the assumption that the light source is far from the object surface. The light source of an endoscope is located at its head, so that it is near the object surface. In this case, the image brightness depends on not only the surface gradients but also the distance from the light source to each point on the surface. To deal with this difficulty, we employ an optical system having a point light source at the projection center, which well approximates the imaging system of the endoscope. In addition, we introduce an equal-distance contour, which is a curve composed of surface points at an equal distance from the light source. Propagating these contours using the image shading, we reconstruct the object surface. We use the level-set method for numerical computation. Experimental results for real medical endoscope images show feasibility of this method.

  172. Determination of sign of gaussian curvature of surface having general reflectance property 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 1351 56-63 1997年

    出版者・発行元: Springer Verlag

    DOI: 10.1007/3-540-63930-6_104  

    ISSN:1611-3349 0302-9743

  173. On the classification of singular points for the global shape from shading problem: A study of the constraints imposed by isophotes 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 1351 48-55 1997年

    出版者・発行元: Springer Verlag

    DOI: 10.1007/3-540-63930-6_103  

    ISSN:1611-3349 0302-9743

  174. Shape reconstruction from an endoscope image by shape-from-shading technique for a point light source at the projection center 査読有り

    K Deguchi, T Okatani

    PROCEEDINGS OF THE IEEE WORKSHOP ON MATHEMATICAL METHODS IN BIOMEDICAL IMAGE ANALYSIS 290-298 1996年

  175. Reconstructing shape from shading with a point light source at the projection center: Shape reconstruction from an endoscope image 査読有り

    Takayuki Okatani, Koichiro Deguchi

    Proceedings - International Conference on Pattern Recognition 1 830-834 1996年

    出版者・発行元: Institute of Electrical and Electronics Engineers Inc.

    DOI: 10.1109/ICPR.1996.546140  

    ISSN:1051-4651

︎全件表示 ︎最初の5件までを表示

MISC 74

  1. ImPACT TRC索状ロボット(細)プラットフォーム:多感覚機能を統合した空気浮上型能動スコープカメラ

    昆陽雅司, 安部祐一, 永野光, 坂東宜昭, 山崎公俊, 岡谷貴之, 糸山克寿, 奥乃博, 田所諭

    システム制御情報学会研究発表講演会講演論文集(CD-ROM) 62nd ROMBUNNO.141‐5 2018年5月16日

    出版者・発行元: システム制御情報学会

  2. 感覚機能統合型能動スコープカメラの改良と瓦礫フィールドへの適用

    山内悠, 安部祐一, 永野光, 昆陽雅司, 坂東宜昭, 山崎公俊, 糸山克寿, 猿渡洋, 岡谷貴之, 奥乃博, 田所諭

    計測自動制御学会システムインテグレーション部門講演会(CD-ROM) 18th ROMBUNNO.1C6‐09 2017年12月20日

  3. 感覚機能統合型能動スコープカメラの開発

    安部 祐一, 田所 諭, 坂東 宜昭, 永野 光, 昆陽 雅司, 山崎 公俊, 糸山 克寿, 猿渡 洋, 岡谷 貴之, 奥乃 博

    ロボティクス・メカトロニクス講演会講演概要集 2017 (0) 1P2-P01 2017年5月9日

    出版者・発行元: 一般社団法人 日本機械学会

    DOI: 10.1299/jsmermd.2017.1p2-p01  

    eISSN: 2424-3124

    詳細を見る 詳細を閉じる

    <p>We had developed an active scope camera: the robot video scope that can move by itself to probe narrow gaps for rescue missions. However, to investigate the interiors of collapsed houses effectively, not only the high mobility but also the sensors, such as vision, audition and haptics should be installed on the robot. According to this, we have developed an active scope camera with the sensory functions and demonstrated these functions in the test field which imitates collapsed houses. In the demonstration, we mainly showed four functions. (1)Detection of collision and it's presentation to the operator, (2)Searching operation assist with the processed vision data, (3)Shape estimation of the robot with sound and IMU data, and (4)Voice enhancement to detect victim's voice.</p>

  4. 深層学習とその画像応用に関する最新動向

    岡谷貴之

    日本画像学会誌 56 (2) 157-162 2017年2月

    出版者・発行元: 日本画像学会

    DOI: 10.11370/isj.56.157  

    ISSN: 1344-4425

  5. 索状ロボットのためのロバストな視覚SLAMの実現

    清水寛太, XU Liang, 岡谷貴之, 安部祐一, 昆陽雅司, 田所諭

    日本機械学会ロボティクス・メカトロニクス講演会講演論文集(CD-ROM) 2017 2017年

    ISSN: 2424-3124

  6. 救助犬に搭載した全方位カメラを用いた視覚オドメトリ

    矢田亮太郎, 清水寛太, 岡谷貴之, 大野和則, 山口竣平, 濱田龍之介, 田所諭

    日本機械学会ロボティクス・メカトロニクス講演会講演論文集(CD-ROM) 2017 2017年

    ISSN: 2424-3124

  7. ImPACT TRC太索状ロボット研究開発の現状

    松野文俊, 伊藤一之, 亀川哲志, 田中基康, 有泉亮, 奥乃博, 大道武生, 鈴木陽介, 岡谷貴之, 多田隈建二郎

    日本ロボット学会学術講演会予稿集(CD-ROM) 34th ROMBUNNO.3C2‐05 2016年9月7日

  8. ディープラーニング

    岡谷貴之

    映像メディア学会誌 68 (6) 466-471 2016年7月30日

    出版者・発行元: 一般社団法人映像情報メディア学会

    DOI: 10.3169/itej.68.466  

    ISSN: 1342-6907

    詳細を見る 詳細を閉じる

    多層ニューラルネットを用いる機械学習の方法(ディープラーニング)が最近注目を集めており,画像や音声の認識などの難問を解決し得る有力な方法として期待されている.本解説では,ニューラルネットの基礎から始め,過学習の克服を可能にし今の研究ブームのきっかけを作ったプレトレーニングと,さらには画像認識で本命視される畳込みネットワークについて述べる.

  9. 画像認識のための深層学習の研究動向 : 畳込みニューラルネットワークとその利用法の発展(<特集>ニューラルネットワーク研究のフロンティア)

    岡谷 貴之, Takayuki Okatani, 東北大学大学院情報科学研究科, Graduate School of Information Sciences Tohoku University

    人工知能 = journal of the Japanese Society for Artificial Intelligence 31 (2) 169-179 2016年3月1日

    出版者・発行元: 人工知能学会 ; 2014-

    ISSN: 2188-2266

  10. 画像認識革命;2.ディープラーニングによる画像認識 -畳込みネットワークの能力と限界-

    岡谷 貴之

    情報処理 56 (7) 634-639 2015年6月15日

    出版者・発行元: 情報処理学会 ; 1960-

    ISSN: 0447-8053

  11. 画像認識の現状認識:ディープラーニングの成功と限界?

    岡谷 貴之

    情報処理学会研究報告. SLP, 音声言語情報処理 2015 (11) 1-1 2015年5月18日

    出版者・発行元: 一般社団法人情報処理学会

    詳細を見る 詳細を閉じる

    一枚の画像からそこに写る物体の名前を答える物体カテゴリ認識は,コンピュータビジョンの長年の課題であったが,ディープラーニングの登場,正確には畳込みニューラルネットの 「再発見」 により大きく進展し,人の視覚と比肩するほどまでになった.一方,その他の多くの画像認識のタスク,例えば特定の物体を画像中で位置決めする 「物体検出」 や,人のポーズ認識,さらには動画像を使った人の行動認識に対しては,物体カテゴリ認識ほどはディープラーニングは成功していない.従来法を超えるという意味では一定の有効性が確認されているものの,これらのタスクでの性能は,人の視覚まではまだ相当の隔たりがある.また,物体カテゴリ認識あるいはその類似タスクに極めて高い精度を発揮する畳込みニューラルネットだが,なぜそんなに高い性能が出るのかの理解が追い付いていない.本講演では,この辺りの現状についてまとめ,今後を占うことを試みたい.

  12. Four-dimensional City Modeling using Vehicular Imagery

    Ken Sakurada, Takayuki Okatani

    情報処理学会研究報告. CVIM, [コンピュータビジョンとイメージメディア] 2015 (35) 1-16 2015年5月11日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    We propose a novel method for four-dimensional city modeling using vehicular imagery. Motivation behind this study is to estimate and visualize damage and recovery process of tsunami-damaged area. To conduct the research, we have been recording the images in tsunami-damaged area periodically driving a car on which omnidirectional camera is mounted. To estimate temporal changes of a wide area using vehicular imagery, there are many challenges to overcome, for example, limitation of camera viewpoint, coverage of vehicular imagery and illumination condition. Furthermore, a 3D model is not always available for every city, and the scene images do often not have sufficient visual features to perform precise registration. In the case of a wide-area disaster, it is computationally prohibitive to reconstruct the three-dimensional structure of entire areas. To overcome these difficulties, we propose 2D, 3D and object-based change detection methods. The 2D method detects scene change from an image pair using visual features of convolutional neural network. The 3D method estimates structural changes of scene from images taken at multiple viewpoints even if there is depth ambiguity of a scene. The object-based method estimates land surface condition of an entire city integrating aerial and street-view imagery, which are taken at vastly different viewpoints. The experimental results show that our methods can accurately and effectively estimate temporal changes of a city.

  13. 深層学習について

    岡谷貴之

    日本ロボット学会誌 33 (2) 92-96 2015年4月15日

    DOI: 10.7210/jrsj.33.92  

  14. ディープラーニングと画像認識 : 基礎と最近の動向 (特集 ニューロサイエンスと数理モデリング)

    岡谷 貴之

    オペレーションズ・リサーチ 60 (4) 198-204 2015年4月

    出版者・発行元: 公益社団法人日本オペレーションズ・リサーチ学会

    ISSN: 0030-3674

    詳細を見る 詳細を閉じる

    ディープラーニングは近年,人工知能の色々な分野で大きな成功を収めつつあり,その高い性能は広く知られるようになった.本稿ではディープラーニングの画像への応用,中でも画像認識に焦点を絞り,そこで欠かせない存在となっている畳込みニューラルネットについて,その技術的な基礎から最近の研究の動向までを概観する.

  15. 畳込みニューラルネットワークを用いたシーン変化の検出

    櫻田 健, 岡谷 貴之

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2015 (2) 1-8 2015年2月27日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿はグリッド特徴を利用して各時刻一枚ずつの画像ペアからシーンの変化を検出する手法を提案する.自動のシーン変化の検出は都市管理や,災害の復旧,復興,減災において有用である.本研究の背景として,車載カメラの画像を用いて津波被災地の街並みの変化を可視化する目的がある.既存のシーン変化の検出手法は,街並みの 3 次元モデルや,異なる時刻のデータ間におけるピクセルレベルの位置合わせを必要とした.そのため,正確な 3 次元モデルが得られない本研究のようなケースには適用することが困難である.さらに,広域の 3 次元モデルの復元や,ピクセルレベルでの位置合わせは膨大な計算量を必要とする.これらの問題を解決するために,本研究では畳込みニューラルネットワークとスーパーピクセルセグメンテーションを統合した新しいシーン変化の検出方法を提案する.提案手法はシーンの 3 次元モデルやピクセルレベルでの位置合わせを必要とせず,計算量を大幅に削減できるため,津波被災地全体のシーン変化の推定を可能とする.

  16. 質感の画像認識 : 研究の状況と今後の課題(質感の計測・認識・提示,災害)

    岡谷 貴之

    電子情報通信学会技術研究報告. MVE, マルチメディア・仮想環境基礎 114 (410) 205-205 2015年1月15日

    出版者・発行元: 一般社団法人電子情報通信学会

  17. 画像認識のための深層学習(<連載解説>Deep Learning(深層学習)〔第4回〕)

    岡谷 貴之, Takayuki Okatani, 東北大学大学院情報科学研究科, Graduate School of Information Sciences Tohoku University

    人工知能学会誌 = Journal of Japanese Society for Artificial Intelligence 28 (6) 962-974 2013年11月1日

    出版者・発行元: 人工知能学会

    ISSN: 0912-8085

  18. ディープラーニングによる特徴学習と画像認識への応用(特別講演,計算解剖学における数理モデルの深化と応用展開)

    岡谷 貴之

    電子情報通信学会技術研究報告. MI, 医用画像 113 (146) 69-69 2013年7月11日

    出版者・発行元: 一般社団法人電子情報通信学会

  19. ディープボルツマンマシンを用いた線画の修復

    吉田 孝弘, 嵯峨 智, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2013 (30) 1-6 2013年5月23日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    ディープラーニングの方法は,画像認識の様々なベンチマークテストにおいてよい結果を残しているが,それらは主に畳込みニューラルネットワークを用いた教師あり学習によるものである.一方,画像を学習対象とした無教師学習,特に全結合型の多層ネットワークを用いたものは,必ずしもベンチマークテストで注目されるような結果は残していない.本論文では,ディープボルツマンマシンを用いた多層ネットワークの無教師学習により,線画の生成モデルを学習し線画の欠損修復を行う方法を述べる.その上で自由度を揃え単層のモデルと比較し,全結合型の多層ネットワークの有効性について比較実験を行う.

  20. ディープラーニング (パターン認識・メディア理解)

    岡谷 貴之, 齋藤 真樹

    電子情報通信学会技術研究報告 : 信学技報 112 (385) 111-127 2013年1月23日

    出版者・発行元: 一般社団法人電子情報通信学会

    ISSN: 0913-5685

    詳細を見る 詳細を閉じる

    ディープラーニングは,多層のニューラルネットワーク(以下NN)を用いる機械学習の方法論である.最近,これに基づく方法群が,画像認識のベンチマークテストで従来の記録を次々に塗り替えた他,音声認識やその他の学習・認識の問題に応用されて同様に高い性能を示すに至り,幅広い関心を集めつつある.本稿では,このディープラーニングについて,画像認識への応用を念頭に,現時点で知られている様々な方法をなるべく網羅的に説明する.具体的には, NNの基本構成から, Convolutional NN, プーリング,局所コントラスト正規化,教師なし学習であるプレトレーニング,オートエンコーダ,スパースな特徴表現を可能にする複数の方法, Restrlcted Boltzmann MachineやDeep Belief Networkなどの生成モデルに基づくディープラーニングの各手法を,それぞれ説明する.またディープラーニングのためのソフトウェアライブラリにも触れる.

  21. 車載カメラで撮影した画像列と市街地地図情報を用いた建築物の出現・消失の推定 (パターン認識・メディア理解)

    手塚 大基, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 : 信学技報 112 (385) 51-57 2013年1月23日

    出版者・発行元: 一般社団法人電子情報通信学会

    ISSN: 0913-5685

    詳細を見る 詳細を閉じる

    本研究では車載カメラで走行しながら市街地を撮影した画像群及び,建築物の輪郭情報が記載された住宅地図を使って,両者の間の不整合を自動的に特定する方法を述べる.例えば,住宅地図上には存在するが,実際には消失した建築物の自動判定か行なえる.具体的には画像群からシーンの3次元復元を行い,得られた3次元情報と地図上の建築物との比較を行い,建築物の存在を確率的に推定する.東日本大震災で被災した市街地を撮影して得た被画像と,被災前の住宅地図を使って実験を行ったので,その結果を報告する.

  22. ロバストバンドル調整を用いた車載ステレオカメラのセルフキャリブレーション (パターン認識・メディア理解)

    深澤 至貴, 岡谷 貴之, 出口 光一郎, 木村 宣隆

    電子情報通信学会技術研究報告 : 信学技報 112 (385) 59-66 2013年1月23日

    出版者・発行元: 一般社団法人電子情報通信学会

    ISSN: 0913-5685

    詳細を見る 詳細を閉じる

    車載ステレオカメラのセルフキャリブレーション,つまり車両が一般道を通常走行する際に得られる画像のみを用いたキャリブレーションの方法を述べる.具体的には,各カメラの内部パラメータおよびレンズ歪がオフラインのキャリブレーションて既知であるとし,ステレオカメラを構成する左右カメラの間の相対姿勢を推定する.この場合,自然のシーンを撮影した左右カメラの画像間でいかに外れ値のない点対応を得られるか,あるいは混入した外れ値を除去できるかが鍵となる.走行中の連続画像間およひ左右カメラの画像間て特徴点を追跡・対応付けてその軌跡を得,それらをすべて使用して完全な3次元復元(SfM)を行い,ロバストバントル調整によって全体最適化をすることで,得た点対応に混入する外れ値を効果的に取り除き,高精度に目的パラメータを推定する.実験結果は,この方法がステレオカメラのキャリブレーションに有効であることを示した.

  23. ロバストバンドル調整を用いた車載ステレオカメラのセルフキャリブレーション

    深澤至貴, 岡谷貴之, 出口光一郎, 木村宣隆

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2013 (10) 1-8 2013年1月16日

    詳細を見る 詳細を閉じる

    車載ステレオカメラのセルフキャリブレーション,つまり車両が一般道を通常走行する際に得られる画像のみを用いたキャリブレーションの方法を述べる.具体的には,各カメラの内部パラメータおよびレンズ歪がオフラインのキャリブレーションで既知であるとし,ステレオカメラを構成する左右カメラの間の相対姿勢を推定する.この場合,自然のシーンを撮影した左右カメラの画像間でいかに外れ値のない点対応を得られるか,あるいは混入した外れ値を除去できるかが鍵となる.走行中の連続画像間および左右カメラの画像間で特徴点を追跡・対応付けてその軌跡を得,それらをすべて使用して完全な3次元復元(SfM)を行い,ロバストバンドル調整によって全体最適化をすることで,得た点対応に混入する外れ値を効果的に取り除き,高精度に目的パラメータを推定する.実験結果は,この方法がステレオカメラのキャリブレーションに有効であることを示した.

  24. 車載カメラで撮影した画像列と市街地地図情報を用いた建築物の出現・消失の推定

    手塚大基, 岡谷貴之, 出口光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2013 (9) 1-7 2013年1月16日

    詳細を見る 詳細を閉じる

    本研究では車載カメラで走行しながら市街地を撮影した画像群及び,建築物の輪郭情報が記載された住宅地図を使って,両者の間の不整合を自動的に特定する方法を述べる.例えば,住宅地図上には存在するが,実際には消失した建築物の自動判定が行なえる.具体的には画像群からシーンの3次元復元を行い,得られた3次元情報と地図上の建築物との比較を行い,建築物の存在を確率的に推定する.東日本大震災で被災した市街地を撮影して得た被画像と,被災前の住宅地図を使って実験を行ったので,その結果を報告する.

  25. 過去の街並みを可視化するスマートフォンを用いた拡張現実

    佐藤慎也, 岡谷貴之, 出口光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2012 (3) 1-4 2012年5月16日

    詳細を見る 詳細を閉じる

    本研究では,身近なスマートフォンを用いて,過去の街並みを可視化する拡張現実(AR)を実現する.ある街の通りでスマートフォンをかざすと,その方向の過去の街並みがスマートフォンの画面上に表示され,あたかもその画面が窓となり,その中に時間をさかのぼった景色が見えるようなものである.画面に表示する画像は,あらかじめ撮影した全方位画像から切り出して生成する.このとき,端末の姿勢を内部センサおよび画像を組み合わせて推定し,リアルタイムで切り出し領域を変えて表示することで,上述の拡張現実感を得られるようにする.端末の姿勢がユーザの見たい方向と一致しない場合でも,端末の表示画面側にあるカメラを用いてユーザの顔を検出し,端末とユーザの顔の相対位置を推定し,この不一致を補正する.ユーザが今いる位置は,端末のGPSデータと,端末のカメラがその場で撮影した画像を両方利用して推定する.具体的には,GPSによっておおまかな位置を推定し,それによって全方位画像のデータベースから対象とする画像集合を絞り込む.その後,撮影画像との間で画像特徴を用いた類似度を求め,位置を高精度に推定するとともに,表示すべき画像を決定する.

  26. 自然画像を対象とするランキング学習に基づく質感認識の一手法

    阿部 厳, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2012 (7) 1-7 2012年3月8日

    詳細を見る 詳細を閉じる

    本研究の目的は,物体表面の光沢感や透明感といった質感に関するいくつかの属性の尺度を画像から推定する方法を実現することである.自然画像を扱えるようにするため,特定の質感に関する属性について,2枚の画像間で人がその強弱を判断した情報を学習データとし,これを使ったランキング学習を行うことで,その属性の強弱を表現する関数を求める手法を提案する.Flickr Material Databaseに含まれる画像を用いて,このような質感に関する属性の比較情報を含むデータセットを生成し,これに提案手法を適用した.属性の強弱を表現した関数を,光沢感,透明感,なめらかさの3つの質感属性について学習し,得られた関数の性能を実験によって評価したところ,良好な結果を得た.さらにこの関数を,金属,ガラス,石といった材質を分類する問題に利用する簡単な実験を行い,その有効性を確認した.The goal of this research is to develop a method for recognizing the quantitative values of several attributes associated with surface qualities of an object, such as glossiness and transparency, from its image. To make it possible to deal with natural images, we propose to use for learning data a set of relative information representing which of a pair of images has a larger value in terms of a particular attribute, which is created by humans, and then to obtain a function representing the relative strengths of the attribute based on learning to rank. We create a data set containing such relative information for several attributes using the images from Flickr Material Database, and then applied our method to the created data set. For three attributes of glossiness, transparency, and smoothness, we learn the functions representing the relative strengths of these attributes. We present a few experimental results in which the learned functions show good performances in predicting relative relations between a given pair of images. We also perform a preliminary experiment in which the learned functions are applied to the classification of materials, such as metals, glasses, and stones. The experimental results show the effectiveness of this approach.

  27. 建物が疎に並ぶ通りの多視点パノラマ画像生成

    柳澤 惇, 手塚大基, 櫻田 健, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2012 (9) 1-8 2012年3月8日

    詳細を見る 詳細を閉じる

    本稿では,市街地の通りのような長いシーン,特に通り沿いに建物がまばらにしか立っていないような場合を対象に,多視点パノラマ画像を生成する方法を述べる.過去の研究で知られているように,そのように奥行きの変化幅が大きいシーンを対象に,ひずみのないパノラマ画像を生成することは原理的に不可能である.この困難さを克服するため,提案方法は,通りに面する建物の表面および地面のみを鮮明に描画する.それ以外の物体はぼかして提示することで,それらの存在感は保ったまま幾何学的な歪みをなるべく目立たなくする.これらを実行するため,最初に対象とするシーンの3次元的構造をSfM(Structure from Motion)によって推定し,シーンの各点が建物表面,地面とそれ以外のどれに属するかを,MRF(Markov Random Field)モデルに基づく最適化により同定し,これを元に,これら3種類の物体の画像をモザイキングすることでパノラマ画像を合成する.画像のぼけは,ライトフィールドフォトグラフィーで使われるのと同様の方法で生成する.2011年3月11日の東日本大震災に伴う津波被災地を対象に,提案方法を用いて生成したいくつかのパノラマ画像を示す.This paper presents a method for creating multi-viewpoint panoramas that is particularly targeted at streets with sparsely located buildings. As is known in the literature, it is impossible to create panoramas of such scenes having a wide range of depths in a distortion-free manner. To overcome this difficulty, our method renders sharp images only for the facades of buildings and the ground surface (e.g., vacant lands and sidewalks) along the target streets; it renders blurry images for other objects in the scene to make their geometric distortion less noticeable while maintaining their presence. To perform these, our method first estimates the three-dimensional structures of the target scenes using the results obtained by SfM (structure from motion), identifies to which category (i.e., the facade surface, the ground surface, or other objects) each scene point belongs based on MRF (Markov Random Field) optimization, and creates panoramic images of the scene by mosaicing the images of the three categories. The blurry images of objects are generated by a similar technique to digital refocus of the light field photography. We present several panoramic images created by our method for streets in the tsunami-devasted areas in the north-eastern Japan coastline because of the Great East Japan Earthquake of March 11, 2011.

  28. 両眼視差と被写界深度ボケを同時に提示する注視反応ディスプレイ

    折笠達郎, 岡谷貴之, 出口光一郎

    画像の認識・理解シンポジウム(MIRU2011)論文集 2011 1201-1208 2011年7月20日

  29. 画像の標本化過程を考慮した高精度かつ頑健な平面の追跡

    伊藤栄介, 岡谷貴之, 出口光一郎

    画像の認識・理解シンポジウム(MIRU2011)論文集 2011 440-447 2011年7月20日

  30. 2A2-O09 拡散光を利用したタッチパネル型反射型触覚センサの設計と実装(触覚と力覚(1))

    小池 桃太郎, 嵯峨 智, 岡谷 貴之, 出口 光一郎

    ロボティクス・メカトロニクス講演会講演概要集 2011 "2A2-O09(1)"-"2A2-O09(4)" 2011年5月26日

    出版者・発行元: 一般社団法人日本機械学会

    詳細を見る 詳細を閉じる

    With the advancement of information technology device, many tactile sensing device are developing. We aim at simple but sensitive tactile sensor and have proposed reflection-based tactile sensor. The reflection-based tactile sensor can employ LED (Light Emitting Diode) and PD (Photo Diode) as thin and fast sensing devices. However this sensing method does not have optical lens. This results in low density sensing because the sensor cannot employ imaging technology. The essence of the sensor is the distribution of brightness caused by diffused light. We have investigated the changes of three dimensional distribution according to simple deformation pattern through simulation. This time we consider the capability of this method from minute simulation, and design the reflection-based tactile sensor employing LED and PD.

  31. 2P1-O14 反射型触覚センサにおける全体最適化を用いたセンサ面形状復元手法の実装(触覚と力覚(2))

    平良 亮祐, 嵯峨 智, 岡谷 貴之, 出口 光一郎

    ロボティクス・メカトロニクス講演会講演概要集 2011 "2P1-O14(1)"-"2P1-O14(4)" 2011年5月26日

    出版者・発行元: 一般社団法人日本機械学会

    詳細を見る 詳細を閉じる

    Because of increasing demands for robots designed to work in the living environment and intuitive interfaces such as iPhones, it is necessary to develop high performance tactile sensors. We are developing a total-internal-reflection based tactile sensor, which reconstructs the sensor surface from a set of feature points on a reflection image of a image pattern. Our proposed method solves Geometric optical constraints by total optimization to reconstruct the sensor surface. By reconstruction from simulated images, we found the proposed method could perform high accurate reconstruction in real time and it was applicable for various sensor shapes and robust to tracking error. Then, we implemented the proposed method on the actual sensor and confirmed the real-time reconstruction.

  32. マーカーとアピアランスを併用した手の姿勢認識

    阿部 厳, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2011 (26) 1-2 2011年5月12日

    詳細を見る 詳細を閉じる

    画像を用いた手の姿勢認識には幅広い応用があり,様々な手法がこれまでに提案されている.最も安定かつ高精度な方法は,マーカーベースのモーションキャプチャシステムであるが,手にたくさんのマーカーを取り付ける必要があり,ユーザの負担が大きい.アピアランスベースの方法は,このような問題はないものの,精度や頑健性の点で大きく劣るか,認識可能な手の姿勢の自由度に制約がある.以上を踏まえて本研究では,少数のマーカーを使って得られる幾何学的な情報と手のアピアランスの情報を組み合わせることで,ユーザの負担と精度をバランスできる方法を述べる.Image-based recognition of human hand poses has a large number of applications and many methods have been proposed so far. The most stable and accurate approach is marker-based motion capture. However, it requires that many markers are attached to the hands, which deteriorates the usability of the system. Appearance-based methods, which are free from such problems, are much inferior in terms of accuracy and robustness, and often have limitations in the degrees of freedom of the hand poses that can be recognized. Based on these considerations, this study presents a method that can balance the system usability and the recognition accuracy, by combining the geometric information obtained by using a small number of markers with the appearance of the hands.

  33. 画像の標本化過程のモデルに基づく高精度・頑健な平面追跡 (パターン認識・メディア理解)

    伊藤 栄介, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 110 (381) 127-134 2011年1月20日

    出版者・発行元: 電子情報通信学会

    ISSN: 0913-5685

  34. 注視反応ディスプレイによる両眼視差とボケの同時提示

    折笠 達郎, 鈴木 孝明, 岡谷 貴之, 出口 光一郎

    コンピュータビジョンとイメージメディア(CVIM) 2011 (40) 1-8 2011年1月13日

    出版者・発行元: 情報処理学会

    ISSN: 1884-0930

    詳細を見る 詳細を閉じる

    眼の被写界深度の制約により眼の網膜上に生成される像のボケは,重要な奥行き手掛かりの一つである.ビューアの注視位置を計測し,これに基づいてリアルタイムに表示映像を変えるやり方で,このボケを映像上に疑似的に再現する映像ディスプレイシステムがある.本稿では,これを眼鏡方式の3次元ディスプレイと組み合わせることで,被写界深度制約によるボケと両眼視差を,ビューアに同時に提示するシステムについて述べる.このシステムを使って,2つの奥行き手掛かりを提示することによる奥行き感や臨場感の強化の度合いを調べる簡単な実験を行ったので,それについても述べる.Retinal image blur generated by the depth-of-field (DOF) limitation of eyes is one of important depth cues. The image display system was proposed that virtually reproduces this image blur by measuring the gaze of the viewer and changing the displayed image in real time based on the gaze measurement. In this paper, we present a new display system that can simultaneously present the DOF blur and binocular disparity by combining the previous system with a 3D display based on the alternate-frame sequencing. We also show a few results of preliminary experiments conducted to examine the magnitude of the enhancement in depth sensation and visual presence by simultaneously presenting both the depth cues.

  35. GPUを使った高速高精度な平面トラッキング

    伊藤 栄介, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2010 (30) 1-5 2010年5月20日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本研究では,Malis らの提案した ESM による画像追跡手法を高速かつ高精度に実現することを目的とし,GPU に実装する方法を述べる.ESM は追跡対象全ての画素を使った計算である為,特徴点がとり難い画像では有効な手法である.ESM は元々高速な計算が可能となるように設計されているが,ESM の並列計算可能な部分に着目し,GPU の特性を生かしたメモリの使い分けや並列計算をすることで,さらに高速なトラッキングを実現した.加えて,テスト画像を用いて既存の手法との計算時間およびトラッキング性能の比較を行った.This paper presents a method of implementing the ESM visual tracker proposed by Malis et al. to realize fast and accurate visual tracking. The ESM tracker is effective especially for the images in which feature points are difficult to obtain, since it uses entire image pixels of the target image region. It is originally designed so as to enable fast tracking, and our GPU implementation enables yet faster tracking; we focus on the parts of the algorithm that can be processed in a parallel fashion, and implement it by appropriately using the memory system of a GPU based on its properties and performing parallel computation. We also show comparison between the GPU and a CPU implementations of the ESM algorithm in terms of computation time and tracking performance.

  36. 幾何学的位置計測手法と照度差ステレオの融合による高精度形状計測

    関 拓郎, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2010 (31) 1-7 2010年5月20日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    物体の形状を計測する方法として,プロジェクタからパタン光を物体に投影したものをカメラで撮影して,カメラとプロジェクタの画像間の視差を利用するものがある.この方法では計測精度は画像の解像度などで決まる視差の精度で制約され,得られる形状には一定の誤差が残る.これと異なる方法として,照明方向を変化させて撮影した画像の濃淡から形状を推定する照度差ステレオがある.この方法は一般的に反射特性や照明条件を事前に知らなければならず,高精度な形状を得ることは難しい.一方で,法線を直接求めるため,隣り合う点の誤差が少ないという利点がある.本研究では,視差を用いて求めた形状の情報と,画像の濃淡から推定した形状の情報を組み合わせて高精度な形状を得る方法を示す.There is a method for measuring the 3D shape of objects by using the parallax between the camera and projector images by taking a picture of the objects onto which striped pattern is projected. The accuracy of this measurement is limited by the accuracy of parallax, which is mostly determined by the image resolution. As a result, there have to be some errors in the measured shape. Another method is to estimate the object shape from its shading obtained when lit from difierent direction. In general, this method requires the knowledge of the reflection property of the surface and the lighting condition, and it is difficult to obtain accurate shapes. On the other hand, the method can directly obtain the surface normal and thus there tends to be small errors between neighboring pixels. This paper shows a hybrid method that can obtain an accurate shape by integrating the information of parallax and shading.

  37. カーネル部分最小二乗分析を用いた歩行者認識

    阿部 厳, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2010 (21) 1-7 2010年3月11日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    歩行者認識の手法として,Shwartz らは Histgrams of Oriented Gradients(HOG) にテクスチャおよび色情報を表現する特徴量を付加し,表現力に富んだ高次元な特徴量を用いる手法を提案した.彼らはその高次元な特徴量を一般的な機械学習の手法を用いて分類するために部分最小二乗分析 (PLS) を用いて次元削減を行い,良好な結果を得た.本稿では次元削減手法に着目して,サポートベクトルマシン等でも利用されているカーネル法を PLS に適用したカーネル部分最小二乗分析 (KPLS) を用いることを検討した.INRIA 人画像データセットを用いて,PLS と KPLS の性能の比較実験を行ったところ,KPLS は PLS に対して優位な性能を示した.Shwartz et al. have recently proposed a method for pedestrian detection that uses a very high-dimensional, discriminative feature obtained by combining HOG descriptors with additional color and texture features. To deal with the high dimensional feature by classical machine learning algorithms, they employed Partial Least Squares (PLS) Analysis, an efficient dimensionality reduction technique, and reported promising results. In this paper, focusing on dimensionality reduction, we examine the possibility of applying Kernel Partial Least Squares (KPLS) Analysis, a variant of PLS that uses the kernel method widely used in other classification/recognition methods such as SVM. We experimentally compare PLS and KPLS in terms of detection accuracy using the INRIA pedestrian dataset. The results show that KPLS outperforms PLS.

  38. 配線自動化のためのワイヤハーネスの画像認識 (パターン認識・メディア理解)

    米田 雅樹, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 109 (374) 59-65 2010年1月21日

    出版者・発行元: 電子情報通信学会

    ISSN: 0913-5685

  39. 配線自動化のためのワイヤハーネスの画像認識 (音声)

    米田 雅樹, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 109 (375) 59-65 2010年1月21日

    出版者・発行元: 電子情報通信学会

    ISSN: 0913-5685

  40. 配線自動化のためのワイヤハーネスの画像認識 (マルチメディア・仮想環境基礎)

    米田 雅樹, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 109 (376) 59-65 2010年1月21日

    出版者・発行元: 電子情報通信学会

    ISSN: 0913-5685

  41. 配線自動化のためのワイヤハーネスの画像認識

    米田 雅樹, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2010 (10) 1-7 2010年1月14日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    自動車の配線に使用されるワイヤハーネスの姿勢を画像上で認識する方法を述べる.ワイヤハーネスは,その柔軟性による姿勢変化の自由度の高さと,画像特徴の乏しさから,画像認識の対象としてはかなり難しい部類に入る.本稿ではこの困難さを解決するため,互いに階層構造を持つ複数の部分にわけてワイヤハーネスを表現し,その階層構造をそのワイヤハーネス自体が持つグラフ構造と緩く対応づける.各部分ごとの画像の類似度と,階層構造が与える各部分の相互関係の制約条件の 2 つを使って,ワイヤハーネスの各部分の画像上の位置を決定する.いくつかの実験結果を通じて,この手法の有効性を示す.This paper presents a method for recognizing the pose of a wire harness in a given image; wire (or cable) harnesses are commonly used in the wiring of automobiles. Wire harnesses can be categorized into a difficult class of objects for visual recognition because of the high degrees of freedom of its pose variation due to its flexibility as well as lack of rich image features. To cope with these difficulties, the proposed method represents a wire harness by a combination of the parts that have a hierarchical structure, which are loosely related to the graph structure of the wire harness itself. The method estimates the position of each part by using the visual similarity of each part as well as the constraint on the positional relation among the parts imposed by the hierarchical structure. We show the efficacy of this method through several experimental results.

  42. 1A1-D07 反射型触覚センサにおける汎用性の高い形状復元手法の提案

    平良 亮祐, 嵯峨 智, 岡谷 貴之, 出口 光一郎

    ロボティクス・メカトロニクス講演会講演概要集 2010 "1A1-D07(1)"-"1A1-D07(4)" 2010年

    出版者・発行元: 一般社団法人日本機械学会

    詳細を見る 詳細を閉じる

    We are developing a reflection type tactile sensor, which is simple and sensitive one. This sensor measures deformation of contact surface from reflection image. In order to solve the inverse problem, we formulated new method of shape reconstruction employing three dimensional vectors. We evaluated our proposed method by reconstructing deformed surface from reflection images produced by simulation.

  43. 携帯型プロジェクタのためのプロジェクタカメラシステムの自動キャリブレーション

    岡谷貴之

    画像ラボ 21 (5) 1-6 2010年1月

    出版者・発行元: 日本工業出版

  44. 多眼カメラシステムによるカメラ運動の高精度実時間推定 (パターン認識・メディア理解)

    武田 悟郎, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告 109 (182) 27-34 2009年8月31日

    出版者・発行元: 電子情報通信学会

    ISSN: 0913-5685

  45. 多眼カメラシステムによるカメラ運動の高精度実時間推定

    武田 悟郎, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2009 (5) 1-8 2009年8月24日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では,カメラを動かして得た画像系列を元に,カメラの運動およびシーンの 3 次元形状を復元する運動からの形状復元 (structure from motion, SFM) において,実時間でこれを行うこと (Realtime SFM) を考える.その成否は,シーンの同一点を画像系列上で追跡した特徴点軌跡を以下に安定して高精度に得られるかにかかっている.本稿では,複数のカメラを剛体結合した多眼カメラシステムを用いて,特徴点軌跡のロバストかつ高精度な計算を行うことを提案する.カメラが運動する中,常に新しい特徴点が画像に発生しまた消えてゆくという最も難しい状況において,実時間性の前提から,現在までの画像系列しか得られない因果性の制約を,効果的に解決できることを示す.This paper considers the problem of realtime structure from motion (SFM), in which the camera motion and the scene structure are recovered in realtime from the image sequence captured by moving the camera. Its success entirely depends on whether or not the loci of feature points are obtained stably and accurately that are the results of tracking identical scene points across the image sequence. In this paper, we propose to use a multi-camera system such that multiple cameras are rigidly combined in order to robustly and accurately obtain the loci of feature points. We show that this method can solve the difficulty with the temporal causality that only the images up to the latest frame are available; as a result, the most difficult cases can be dealt with where new feature points always appear and some of the current points always disappear while the camera moves.

  46. モバイルプロジェクタのためのプロジェクタカメラシステムのオンライン自動校正

    岡谷 貴之, 阿部 大輔, 琢磨 直弘, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2009 (4) 1-8 2009年8月24日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では,モバイルプロジェクタから平面スクリーンに像を投影し,その像を固定カメラが常時捉えるような構成のプロジェクタカメラシステムに対する校正法を述べる.応用の1つは,手で持った小型プロジェクタでスクリーンに像を投影するとき,プロジェクタの姿勢によらず常にスクリーンの決まった場所に,幾何学的に補正された画像が表示されるようなシステムである.提案方法は,投影面上の(人工)マーカーを前提とせず,校正用の特別な画像を投影するなどといった,オフラインでの校正手続きなしに,自動的にシステムを校正する.このため,ユーザは見たい映像を投影しながらただプロジェクタを動かすだけでよい.この間,プロジェクタからスクリーン上に投影される画像と,これをカメラで撮影した画像のペアから,カメラ画像とスクリーン間の射影変換がオンラインで推定される.このアプローチは,投影面上のマーカーならびに校正専用の画像を使わないことから,高精度を確保するのが難しい.これを解決するため,逐次バンドル調整に基づく校正方法を示す.この方法は,オンライン校正の性質,すなわち観測データを時々刻々蓄積でき,故に大量のデータが利用可能となることを利用して,計算量と記憶域を最小化しつつ高精度の校正を実現する.This paper proposes a calibration method for a projector-camera system consisting of a mobile projector and a stationary camera. As a target application, we consider a system in which rectified images are always displayed in a fixed position of a planar screen, no matter how the projector is moved. The proposed method automatically calibrate the system without using fiducials on the screen or a special calibration procedure in which dedicated images for calibration are used. Owing to this automatic nature, the user needs only to arbitrarily move the projector by his/her hand. While the user is moving the projector, from the pairs of an image to be projected from the projector and the image of the projected image on the screen that is captured by the stationary camera, the projective transformation between the camera and the screen is estimated in an online manner.

  47. 被写界深度ボケと運動視差を同時に再現する注視反応ディスプレイ

    鈴木 孝明, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2009 (15) 1-8 2009年6月2日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    われわれが 3 次元的シーンを見る際,見つめる対象の奥行きに応じて,眼は半ば自動的に焦点調節を行う.このとき,眼の網膜像には,被写界深度の制約からくるデフォーカスボケが存在するが,これが奥行き手がかりとなることは良く知られている.石澤らは,通常の映像ディスプレイと視線計測装置を使って,人の眼の焦点調節をシミュレートするシステムを提案し,空間情報の視覚提示の新しい方法の実現を狙った.本研究では,これを一歩進め,運動視差とボケを同時に提示するシステムを実現した.立体・運動視差は,ボケよりも強い奥行き手がかりとして知られるが,双方は互いに相補的であり,同時に提示することで相乗効果が見込めるのではないかという期待がある.システムの実装と,行った実験について報告する.When we see a three-dimensional scene, our eyes automatically accomodate (i.e. adjust their focus) depending on the depth to objects that we look at. There is defocus blur in its retinal image that emerges due to the depth-of-focus limitation, which is well known to serve as a depth cue. Ishizawa et al. proposed a system that uses an ordinary image display and an eye tracking system to simulate the accommodation of human eyes using, which aimed at realizing a novel method of visually presenting scene depth information. In this study, extending their study, we realize a system that simultaneously presents motion disparity and defocus blur. Binocular and motion disparity, which is known to be stronger depth cue than defocus blur, is also known to be complementary with defocus blur. Thus, their simultaneous presentation to a viewer is expected to bring about a synergetic effect. We report the implementation of the system and the results of a few experiments we conducted.

  48. バンドルアジャストメント

    岡谷 貴之

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2009 (37) 1-16 2009年6月2日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    バンドルアジャストメントは,画像から幾何学的なモデルのパラメータを推定する問題において,高い推定精度の達成を目指して,非線形最適化を数値的に-それが大規模なものになっても避けることなく-実行する方法である.本稿では,このバンドルアジャストメントを解説する.具体的には,数値計算の基本的アルゴリズムと実装方法,統計的推測としての理論的背景,そして座標選択に関わるゲージの自由度についてそれぞれ述べる.また,時系列の観測に対するオンライン推定を実現するための再帰的計算方法についても述べる.Bundle adjustment is a general method for the problem of estimating parameters of a geometric model from images. It is characterized by that nonlinear optimization is numerically performed to attain high estimation accuracy, even if the size of the optimization is very large. In this article I present a tutorial on this method. Specifically, I describe basic numerical algorithms with their implementation techniques, theoretical grounds in terms of statistical inference, and gauge freedom associated with the choice of a coordinate system. Moreover, I present a method for recursive computation that realizes online estimation for time series observations.

  49. 携帯プロジェクタのための映像キャリブレーションの一手法

    琢磨 直広, 岡谷 貴之, 出口 光一郎

    研究報告コンピュータビジョンとイメージメディア(CVIM) 2009 (13) 1-7 2009年6月2日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では,携帯プロジェクタのためのプロジェクタカメラシステムの校正法を述べる.システムは,プロジェクタの他,一台のカメラおよび平面スクリーン (投影面) からなる.校正の目的は,携帯プロジェクタを手で持って動かしたとき,その位置・姿勢によらずスクリーン上では常に正しく正規化された画像が投影されるようにすることである.そのためには,スクリーンとカメラの幾何学的関係を知る必要があり,その推定が校正の実体である.提案する校正法は,スクリーン上に人工的なマーカの設置を要せず,プロジェクタを手で動かしながら自然画像を投影する間に,自動的にこの関係を推定する.校正後,画像を高速に補正する方法も述べる.This paper presents a method for calibrating a projector-camera system that is specially targeted at the use with a mobile projector. The system consists of the projector, a stationary camera and a planar screen surface onto which images are projected. The goal of the calibration is to make it possible that the image projected from a hand-held mobile projector be always correctly rectified regardless of the pose of the projector. To do this, it is necessary to estimate the geometric relation between the screen surface and the camera, which is the purpose of the calibration. The proposed method can automatically perform this, while a natural image is projected from the projector moved in an arbitrary manner by a hand; it does not require fiducials on the screen surface. We also present a method for rectifying the projected image in real time after the calibration.

  50. 1A1-D01 移動ロボットのための物体認識と自己位置推定における視覚情報処理

    阿部 広紀, 嵯峨 智, 岡谷 貴之, 出口 光一郎

    ロボティクス・メカトロニクス講演会講演概要集 2009 "1A1-D01(1)"-"1A1-D01(4)" 2009年5月25日

    出版者・発行元: 一般社団法人日本機械学会

    詳細を見る 詳細を閉じる

    When a robot manipulates some objects, the object should be specified by the robot vision. This paper describes a novel object recognition approach which combines a method for generic visual categorization with three- dimensional segmentation method. The approach is as follows; first, the object area in the image is cut out by clustering 3D map of features obtained by SLAM. Next, by using the generic visual categorization method, "Bag-of-Keypoints," to the area, robot can recognize what it is and where it is in the image. Our approach contributes to the object recognition that can correspond to the diversity of objects and the image segmentation that uses 3D positions of features which compose the object. In this paper, we discuss the effectiveness of our approach through some tableware detection experiments.

  51. 多眼カメラを使った事前学習を要しないアピアランスベース姿勢推定

    武田 悟郎, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2008 (36) 93-98 2008年5月1日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    画像から物体の姿勢をアピアランスベースで推定する従来の方法は,あらかじめ得た画像と姿勢との学習サンプルを元に推定を行うというものであった.したがって,物体の姿勢の自由度が増すにつれ学習が大変になる(特にサンプルの取得に要する時間)ことや,そもそも学習した物体以外ではうまく推定を行うことができない問題があった.少数のカメラを組み合わせて使う Yang らの微分カメラは,これらの問題を解決すべく提案されたものだが,姿勢変化とアピアランスの関係が線形近似できるという仮定に基づいている.両者の関係は本来かなり非線形なものであって,その齟齬からくる手法の限界(推定が行えるシーンの種類等)は明白である.本稿では,微分カメラよりも多量のカメラを使って,線形近似によることなく,直接姿勢変化を推定する方法について述べる.具体的には,基準となるカメラのある時刻の画像と他のカメラの後の時刻の画像とを,射影変換を作用させながら比較することで,物体の姿勢変化を推定する.Conventional appearance-based methods for estimating the pose of an object from its image use the learning samples of object appearance under known poses that are acquired in advance. These methods suffer from the increase in the time required for acquring the learning samples with the increase of the degrees of freedom of the pose parameters. Furthermore, these methods cannot deal with unknown objects other than the ones that have been learned in advance. The differential camera proposed by Yang et al., which was developed to solve these problems by using a rigid combination of a few cameras, entirely depends on the assumption that the relation between the pose change and the appearance change can be well approximated by a linear function. However, the relation is in essence highly nonlinear, which clearly leads to the limitations of the method. In this paper, we propose a method that uses a larger number of cameras than the differential camera and directly estimates the object pose without the above linearity assumption. To be specific, the method estimates the changes of the object pose by comparing the image captured at a time by a refenrece camera with the images that are captured later by other cameras and are warped by projective transformations.

  52. 複数プロジェクトの重ね合わせ投影による高精細画像実現

    和田 幹生, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2008 (27) 313-320 2008年3月11日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    最近,プロジェクタを複数台使いそれらの投影像を重ね合わせることで,単体のプロジェクタを超える解像度をもつ画像が表示できることが示された.われわれは以前,この方法が実現する画像の画質に関するいくつかの特徴的な性質を,理論的な解析に基づいて示した.具体的には,実現画像の画質が,表示させようとする目標画像の濃淡構造や全体的な明るさに依存することなどである.本稿では,実際にシステムを構築しそれを用いた実験を行って,それらが現実のシステムでも成立することを示す.さらに,重ね合わせた投影のためのシステムのキャリブレーション方法を示す.これは,3 次元形状復元で一般的な位相ソフト法を本システムに応用したもので,精度において従来法を大きく上回ることを示す.Recently, it has been shown that it is possible to realize a higher-resolution image by superimposing multiple low-resolution images projected by different projectors. In our previous study, we pointed out several peculiar properties with the images realized by this superimposed projection based on theoretical analyses. For example, the quality of the realized images depends on the brightness structure and the overall brightness of a target image. In this paper, we show that they hold in the case of real systems by constructing a working system of superimposed projection and conducting experiments. Furthermore, we present a novel method for the calibration of the system, which is based on phase shifting that is widely used in 3D shape reconstruction. We then show that the method is superior to conventional methods in terms of calibration accuracy.

  53. 複数プロジェクタの重ね合わせ投影による高精細画像実現

    和田 幹生, 岡谷 貴之, 出口 光一郎

    電子情報通信学会技術研究報告. PRMU, パターン認識・メディア理解 107 (539) 313-320 2008年3月10日

    出版者・発行元: 一般社団法人電子情報通信学会

    ISSN: 0913-5685

    詳細を見る 詳細を閉じる

    最近,プロジェクタを複数台使いそれらの投影像を重ね合わせることで,単体のプロジェクタを超える解像度を持つ画像を表示できることが示された.われわれは以前,この方法が実現する画像の画質に関するいくつかの特徴的な性質を,理論的な解析に基づいて示した.具体的には,実現画像の画質が,表示させようとする目標画像の濃淡構造や全体的な明るさに依存することなどである.本稿では,実際にシステムを構築しそれを用いた実験を行って,それらが現実のシステムでも成立することを示す.さらに,重ね合わせ投影のためのシステムのキャリブレーション方法を示す.これは,3次元形状復元で一般的な位相シフト法を本システムに応用したもので,精度において従来法を大きく上回ることを示す.

  54. プロジェクタ投影像の超解像化:複数プロジェクタ投影像の重ね合わせによる高解像度画像表示

    岡谷貴之

    画像ラボ 19 (11) 65-69 2008年1月

    出版者・発行元: 日本工業出版

  55. プロジェクタ・カメラシステムを用いた仮想反射特性の生成 (第24回センシングフォーラム資料--センシング技術の新たな展開と融合) -- (パターン計測(3))

    田中 文武, 岡谷 貴之, 出口 光一郎

    センシングフォ-ラム資料 24 313-318 2007年10月25日

    出版者・発行元: 〔計測自動制御学会〕

    ISSN: 1343-7631

  56. アクティブライティングによる画像からの対象物の抽出

    岡崎 智也, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2007 (42) 113-120 2007年5月15日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では、静的な環境を対象に、プロジェクタを使って能動的に照明を制御することで、画像中の対象物体を正確に背景と分離し切り出す方法を述べる。提案方法は、カメラとプロジェクタが作るステレオ対における、対象物体と背景との奥行きの差に起因する視差を利用する。背景のみのシーンと対象物体が存在するシーンのそれぞれに同じパタン光を投影し、そのとき取得した 2枚の画像の輝度値の変化から対象物体を切り出す。通常の「受動的な」背景差分法では扱いが不可能な、対象物体と背景の見えに差がない場合でも、提案手法は視差に基づいて確実に分離を行える。加えて、視差が見えの変化を生成しない領域を複数のパタンを用いて処理する方法や、対象物体の影を複数方向の照明を用いて処理する方法を示す。実験を通じて、これらの方法の有効性を示す。This paper presents a method for segmenting a target object from the background of a scene on its image by actively controlling lighting using projectors. The method uses the image disparity between the target object and the background scene due to the difference in their depth that emerges on the stereo pair of the camera and a projector. An identical pattern is projected onto the background scene and the scene with the target object, respectively, and two images are captured for them. Then, the image region of the target object is identified using the difference in the brightness between the two images. When there is no difference in the natural appearances of the target object and the background scene, it is impossible for the ordinary "passive" background subtraction method to segment them, whereas the proposed method can accurately segment them based on their disparity. We further present a method that uses multiple patterns to deal with the image regions such that the disparity does not result in appearance changes, and also present a method that uses multiple projectors to cope with shadows generated by the target objects. The efficacy of these methods is shown through several experiments.

  57. 非観測データを伴う行列の因子分解のための数値解法について

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2005 (88) 123-130 2005年9月6日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では,非観測成分を含む場合の行列の因子分解(主成分分析)の問題を考える.この問題で行列の分解が一意に定まるかどうかは,観測成分の観測行列内の分布の仕方に依存する.この分布についてわれわれは,分解が一意に定まるための必要十分条件を導く.さらに,この間題のための最も基本的なアルゴリズムであると考えられる,Wibergのアルゴリズムを改めて評価し直す.同アルゴリズムの導出過程と適切な実装方法を示し,上述の一意性条件との関係について述べる.また,Wibergのアルゴリズムは,計算量と収束性能の面で他のアルゴリズムよりも優れていることを実験結果を交えて述べる.This paper considers the problem of factorizing a matrix (or principal component analysis) with missing components. Whether a matrix is uniquely factorized or not depends on how the missing components distribute in the matrix. In this paper, we derive a sufEcient and necessary condition on the missing components distribution for unique factorization. We also review the Wiberg algorithm, supposedly the most basic algorithm for the problem. We summarize the derivation of the algorithm and describe its appropriate numerical implementation, including its relation to the above uniqueness condition. We also show through experimental results that the Wiberg algorithm shows better performance than other algorithms with respect to computational cost and convergence property.

  58. カメラキャリブレーション

    植芝俊夫, 岡谷貴之, 佐藤智和

    情報処理学会コンピュータビジョンとイメージメディア研究報告 148 (18) 1-18 2005年3月3日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    カメラキャリブレーションについて,特にその幾何学的側面に絞って最近の主要な研究のサーベイを行う.準備として,本稿で必要とされる射影幾何の基礎事項の解説の後,透視投影カメラモデルを導入してその射影空間における振る舞いについて述べる.そして,このモデルに基づくキャリブレーション方法として,参照物体を利用する方法とそれを利用しないセルフキャリブレーションに分類して様々なアルゴリズムを紹介する.さらにバンドル調整による推定パラメータの高精度化に言及し,最後に透視投影に従わないレンズ歪みのモデル化とその校正について述べる.This article gives a survey of recent studies of geometric camera calibration techniques. After some explanations of fundamentals of projective geometry, a perspective camera model will be introduced followed by a discussion about its behavior in the three-imensional projective space. Next various camera calibration algorithms based on this model are presented which are classified into two categories: methods using reference objects and self-calibration techniques without them. Then the refinement of estimation parameters through bundle adjustment will be explained and finally modeling and calibrating the effect of lens distortions not obeying perspective projections will be discussed.

  59. メカニカルセンサとイメージセンサの融合による3次元形状復元

    岡谷 貴之

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2005 (4) 123-130 2005年1月21日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本稿では,角速度センサや加速度センサといったメカニカルセンサをカメラに取り付け,カメラの運動に関する情報を計測することで,画像からの3次元復元などの問題において,カメラだけでは従来難しかったような問題を解決する方法について述べる.特に,難しいとされる特徴点の対応付けの問題が,姿勢を計測するメカニカルセンサを使ってどのように解決できるか,センサの種類の使い分けとその精度についての考察も含めつつ議論する.また,このようなシステムで問題となる,カメラとメカニカルセンサの内部的な座標表現の違いを計算するキャリブレーション方法についても述べる.This paper presents a method that resolves several difficulties with the problems of 3D recovery from multiple images by attaching mechanical sensors, such as gyrosensors and accelerometers, to the camera and then measuring the pose of the camera from these sensors. To be specific, it is discussed how the problems of image matching, which are still difficult to perform, could be solved by introducing mechanical sensors measuring 3D pose of the camera, including discussion as to choice of mechanical sensors and their measuring accuracy. Moreover, several methods are described for calibration of the system, which is to determine the difference in internal coordinate representation between the camera and the mechanical sensors, and which will be a problem need to be carefully considered.

  60. 画像上の人物追跡のための色空間における肌色領域の分割抽出法 (第20回センシングフォーラム 資料--センシング技術の新たな展開と融合) -- (セッション1C2 人物と顔のパターン計測(2)--パターン計測部会企画)

    伊茂治 公介, 岡谷 貴之, 出口 光一郎

    センシングフォ-ラム資料 20 289-294 2003年9月16日

    出版者・発行元: 〔計測自動制御学会〕

    ISSN: 1343-7631

  61. 眼底画像の表示法の研究

    島原 達也, 中島 平, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2003 (41) 31-36 2003年5月8日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本論文では二枚の眼底画像を重ね合わせて一枚の画像を生成する方法について述べる.眼底画像とは眼球底部の領域を撮影した画像である.この眼底画像は瞳を通して撮影されるので一度に撮影できる範囲が限られてしまう.したがって広範囲の眼底領域の画像が必要な場合,複数枚にまたがって撮影される.そこでこのように複数枚にまたがって撮影された眼底画像を一枚の画像に重ね合わせることは診断の際に有用であると言える.ここではAli Can等の提案している眼底画像の重ね合わせの方法を用いて重ね合わせを行う.この方法は特徴抽出から変換まですべて自動で行うことが可能であり,しかもある程度以上重なる領域が存在すればほぼ完璧に重ね合わせることができるという利点を持つ.しかし重なる領域が少ない時には対応点を十分求めることができずに重ね合わせることができるはずの眼底画像の組も重ね合わせに失敗することがあるという欠点も持っている.そこで重なる領域が狭い場合でもできるかぎり特徴点同士の対応がとれるようにこの手法を改良した.This paper presents a method for registering fundus images. The fundus image is the photograph of inner bottom of an eyeball. Because it is taken through the pupil and the view field is restricted to small region,it is useful to register multiple fundus images in diagnosing. In this paper we register the fundus images employing the Ali Can's algorithm. This algorithm has the advantage of registering automatically and accurately when the overlap is wide. But when the region of overlap is narrow, we often fail because of the lack of the correspondences. We discuss the increase of the member of correspondences when the overlap is narrow and propose a modified method to obtain more accurate registration.

  62. 平均値シフト法を用いた複数物体追跡の研究

    川中 大樹, 中島 平, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2003 (41) 9-14 2003年5月8日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本論文では画像中の複数物体をリアルタイムで追跡する手法を提案する.背景差分法を用いて対象物体を安定して画像から抽出できる場合には,得られた対象領域に対して平均値シフト法を用いることで,極めて少ない計算量ながらかなりロバストな対象の追跡が可能である.しかし,画像中に複数の移動物体が存在し,これらがたびたび交差するときには対象物体の区別が必要となる.そこで,オクルージョン発生時に対象領域中の色の分布に応じて確率的な重み付けを施した平均値シフト法を行うことで,対象物体を区別して追跡できるようにした.この重みはオクルージョンが起こったときに追跡対象の色のヒストグラムと重なった対象の色のヒストグラムと比較して作成する.ヒストグラムはオクルージョンが起こっていないときに求めて更新するので,予め対象人物の色情報を知る必要がない.This paper presents a method for tracking mulitple objects in an image sequence in a real time. If the objects can be accurately extracted from the images by background subtraction, it is possible to robustly track the objects with a small amount of computational complexity using the mean shift algorithm. When there are mulitple moving objects in the images and they frequently overlap with each other, it is necessary to identify each of them correctly to track them. An algorithm for this problem is shown which is based on mean shift algorithm with probabilistic weights that are computed from color distribution of each target. The weights are generated at every time when occlusion occurs and they are computed from comparison between the histogram of the target window and those of the overlapped windows. The histograms are periodically updated when there is no occlusion.

  63. 3次元環境理解のための多視点カメラのキャリブレーション

    出口 光一郎, 岡谷 貴之

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2002 (2) 1-8 2002年1月17日

    出版者・発行元: 情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    多視点のカメラからの画像情報を有効に統合するためには,それぞれのカメラの特性と位置姿勢関係が分っている必要がある.動的な3次元環境理解においては,キャリブレーション用に形状が正確に分っている対象物体を用意し,その物体が画像上にどのような像を結ぶかで,カメラのパラメータを同定するというような古典的なキャリブレーションはそのままでは使えない.この応用においては,キャリブレーション用の物体はいつも使えるわけでなく,また,対象の大きさや動きに応じて,カメラのズームや姿勢を変えたい.そこで,対象の3次元形状もカメラのパラメータと共に同定する,セルフキャリブレーションを行なわなければならない.本稿では,このような応用のための多視点カメラのセルフキャリブレーションの枠組みについて報告し,実験結果と共に,問題点についても議論する.To understand a dynamic 3D scene by multiple cameras, thecharacteristics as well as locations and poses of the cameras must beidentified. For such the dynamic scene understandings, the conventional camera calibration techniques using some specific calibration objects has a limitation. In the application here, the object shapes are unknown, and some camera parameters such as focal length and camera pose like to be variable. Then, so-called the self-calibration must be employed, where the camera parameters and the object shapes are identified simultaneously and dynamically. In this paper, we introduce our strategy of implementation for the calibration for multi-view cameras based on the self-calibration, show the experimental results, and discuss the problems concerning the dynamic calibration of the multi-view cameras.

  64. 画像からのカメラの姿勢 3姿勢形状復元における推定精度の限度の限界について

    岡谷貴之

    第6回画像の認識 理解シンポジウム講演論文集 335-340 2002年

  65. 見え方に基づく線形モデルあてはめによる画像からの姿勢推定

    出口 光一郎, 岡谷 貴之

    センシングフォーラム資料 18 19-24 2001年10月3日

    出版者・発行元: 〔計測自動制御学会〕

    ISSN: 1343-7631

  66. 3次元向きセンサを取付けたカメラを用いた投票によるカメラの並進運動の推定

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告ヒューマンコンピュータインタラクション(HCI) 2001 (87) 25-32 2001年9月13日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    カメラを自由に動かして撮影したシーンや物体の2枚の画像に対して,画像内の点の対応が未知の場合に,カメラの位置と姿勢を推定する問題について述べる.カメラの姿勢は,カメラに3次元向きセンサを取付けてそこから得る.カメラの内部パラメータを既知とし,カメラの位置すなわち視点間の並進運動だけを,画像とセンサから得た姿勢情報から推定する.このカメラの並進は,対応の可能性がある2点の組を投票することにより推定される.これは点の対応関係を陽に決めずに行われる.この種の問題には,従来よりランダムサンプリングに基づく手法が適用され,その有効性が知られている.提案する投票による推定方法と頑健さについて比較する.This paper is concerned with the problem of estimating pose of a camera in space from a pair of images of a scene or an object that is taken with the camera in a case where the point correspondences between the images are unknown. The camera orientation is obtained by a 3D orientation sensor rigidly attached to the camera. Assuming that the intrinsic parameters of the camera are known, we estimate only the camera position, i.e. translation between viewpoints, from the images and the camera orientation obtained from the sensor. This camera translation is estimated by voting pairs of points that likely correspond to each other. This is done without explicitly determining the point correspondences. Several methods using random samples have been applied to these kinds of problems and their effectiveness has been already known. We compare these methods with our voting-based method in terms of robustness.

  67. 固有空間法はなぜうまく働くか

    出口 光一郎, 岡谷 貴之

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2001 (66) 1-8 2001年7月5日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    最近、画像解析に広く固有空間法が応用され、実用的な手法としての立場を確立している。しかし、画像そのものをデータとして見たときには、サンプル数に比べて次元数が遥かに大きいという、本来の固有空間法を適用するデータとは違った様相を持つ。ここでは、このようなデータに固有空間法を適用することで、副次的な効果としての次元数を下げる、ということが画像解析におけるモデルあてはめに本質的な役割をしているのではないかと論じる。ここでは、画像からの姿勢推定というタスクを例に、画像データそのものに、直接的に姿勢パラメータを結びつける手法をまず述べる。続いて、同じことを、固有空間法を用いて実現し、その対比をすることでこのようなデータの次元数がサンプル数よりはるかに大きいときに何が起きているのかを見てみる。そして、いわゆる「過学習の理論」の立場からこの対比を解釈してみる。The eigen space method has established its position among computer vision techniques, recently, as a tool for robust image analysis as well as efficient image data compression method. However, image data itself has a significant difference from statistical data to which the conventional eigen space method is conceptually applicable, because its dimensionality is much larger than the number of samples which are used to construct the eigen space. Here, we claim that the reduction of the dimensionality which is carried out in the process of the eigen space method plays an improtant role for model fitting to the image data. In this article, we introduce the image based pose estimation task as an examle to apply the eigen space method. We make comparison between the eigen space method and the direct imterpretation method where a linear model combining poses and images. Then, we discuss the comparison from the viewpoint of the theory over-learning.

  68. 多視点動画像からの実時間3次元形状復元の研究

    武田 淳, 岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2001 (36) 75-80 2001年5月10日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    3次元コンピュータグラフィックスを作成するための、画像による動物体3次元形状の測定(いわゆる、モーションキャプチャ)システムの構成について報告する。対象のまわりを取り囲んで配置した多数大のカメラによる画像から、背景を引き去った後、共通部分を抽出するvolume intersection(または、shape from silhouette)の手法は、原理も分かりやすく、高速化による実時間計測の可能性が高い。ここでは、画像から動物体を安定に形状復元することを目的とし、6台のデジタルビデオカメラと1台のPCを用いた実時間形状復元システムの開発を行なった。本報告ではまず始めに,複数の視点から撮影した画像から,実時間で形状を復元する手法について述べる。次に、実際の画像を逆投影して重ね合わせることにより3次元形状を得る際の、実時間で行うための高速化について述べる。現在の構成では、180cm立法の空間を1cmの分解能で形状復元をするとき、3?4フレーム/毎秒の処理速度であるが、同じ構成で、空間の分解能が2cm程度とすれば、毎秒10フレーム以上の復元が可能である。This article reports a configuration of a system for real-time 3D reconstruction from multiple motion images for 3D motion capturing. We employ the volume intersection(or, so called the shape from silhouette) technique. In this system, multiple images are obtained by cameras arranged around a moving object. From the respective images back-ground images are subtracted, then, the object images are back-projected into 3D space. Finally, the volume intersection of each pojection cones is extracted. This technique has great advantages of the simple and high speed processing. Our system has 6 digital camera and one PC, and reconstruct 3D object in (180cm)3 space with 1cm resolution at 3 or 4 frames/s speed. It is possible to speed-up to over 10 frames/s when the resolution is reduced to 2cm. We report the techniques to achieve the processing capability as well as the total configuration of the system.

  69. 多視点画像を用いた画像マッチングと立体形状復元の研究

    人見 康宣, 高橋 克博, 岡谷貫之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 2001 (36) 81-86 2001年5月10日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    キャリプレーションされていないカメラによる複数の画像から、対応点を決定し、対象の三次元形状を復元するシステムの構築について報告する。各画像には座標が既知である点が6点以上あるものとし,それらの点を元に、まず、カメラの相対的な位置,姿勢を計算する。そして、その他の点について、そのカメラ位置、姿勢よりステレオの原理に基づいて、三次元座標を計算する。このとき、各画像間で,同じ対象の点の組つまり画像間の対応を得る必要がある。その作業を自動化する手法がZhang(1994)より提案されている。ここでは基本的にZhangの手法によって対応点の検出を行う。ただし,Zhangの方法では,対応点の候補を絞る段階で緩和法と称しているプロセスについての記憶が明確ではなく、対応点の効率的な絞込みができていなかった。そこで、我々は対応点の候補で特徴点が重複したものを除去するプロセスを作成し,この部分に追加した。これにより対応点を検出する作業を効率化できただけでなく,エピポーラ幾何により正確な対応点を得ることができる。この対応点検出により、大型建造物の空撮画像からの3次元形状復元の結果も合わせて示す。This article reports a technique for extracting correspondences from multiple images to reconstruct 3D shape of the object. The images are obtained by uncalibrated cameras, but assumed to have 6 or more points whose world coordinates are known. From those points, first, relative poses and positions of the stereo principle. To establish these correspondences, we employ the technique proposed by Zhang(1994). But, in his method, the process called as the relacation was not clearly defined and did not efficiently exclude fault correspondences. We modified this process and achieved an effcient and accurate matching. We show total process of image matching based on the epipolar geometry, and also show the results of 3D reconstruction or large buildings from a set of acrial images.

  70. 陰影画像の等濃度線の分布を利用した形状の凹凸の推定

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 1998 (5) 49-55 1998年1月22日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    陰影画像が与えられたとき,特異点(面の向きと光源の向きが一致する最も明るい画像の点)の種類を,極大点,極小点,鞍点のいずれであるかを特定する問題を,特に光源方向が未知である場合について考える.特異点の種類を特定することは,大域的なShape from Shading問題の鍵であり,形状の凹凸情報を得ることと等価である.画像の濃淡の鞍点が,常に形状の放物的曲線(ガウス曲率ゼロの曲線)の上にあるというホトメトリック不変性をもとに,この点からの濃淡の最急上昇曲線によってつながれた二つの特異点が,互いに異なる符号を持つことを示す.微分方程式の数値計算を伴うことなく,画像から直接特異点に関する情報を得られる.This paper is concerned with the problem of identifying types of image singular point as maxima, minima, or saddle points, especially for the case of unknown lighting direction. (Singular points are defined as maximally bright points in the image whose corresponding surface normals coincide with the lighting direction). The identification of singular points is the key to the global shape from shading problem, and it is equivalent to determine convexity/concavity of the object shape. Using photometric invariance that saddle points of the image grey level profile always lie on the parabolic curves (curves of zero Gaussian curvature), it is shown that the two singular points which are connected by two steepest ascent curves starting from an image saddle point have different signs with each other. This constraints the combination of the types of the singular points in the global shape from shading problem.

  71. 等濃度曲線を利用するGlobal Shape from Shadingのための特異点分類の一手法

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 1997 (70) 187-194 1997年7月24日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    なめらかな物体の表面を,画像の陰影から復元するには,画像内の特異点すべてについて,それに対応する形状が凸,凹,鞍点のうちどれであるかを定める必要がある.なめらかな曲面と,それから生じる陰影との間の,微分幾何学的な関係を調べると,濃淡値が等しいような画像上の曲線のトポロジーから,特異点の種類に関する情報を直接引き出すことができる.これを利用すると,これまでのように特異点の分類のために微分方程式(画像照度方程式)を数値的に解く必要がなくなる場合がある.このとき,特異点の分類が確かなものとなり,より正確な形状復元が可能になる.It has been shown that the shape of a smooth object can be globally reconstructed from one shaded image if singular points in the image are correctly classified into minima, maxima and saddle points. This paper shows that the topology of isophotes (lines of equal brightness) around the singular points puts some constraints on the classification. They come from the relation in terms of differential geometry between a smooth surface and its shaded image. In some cases, the numerical solution of a PDE (the image irradiance equation) that the conventional methods required is no longer necessary. This makes it possible to accurately classify singular points and to reliably reconstruct the surface shape.

  72. 光源の方向を変えて得られる陰影画像列からの対象表面の曲率特徴の抽出と領域分割

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 1997 (31) 15-22 1997年3月27日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    本論文では,照明方向を変化させて得られる3枚の画像の陰影から,なめらかな物体表面のガウス曲率の符号を決定する方法について述べる.ランバート反射を仮定すると,照明の正確な方向やその強度,物体表面の反射係数の知識がなくても符号を決定できる.この理由を,面の法線ベクトル分布とガウス写像,さらに画像との間の関係を用いて述べる.提案する方法は,以前の方法では必要だったような照明の方向に対する条件を必要としない.さらに,表面の反射特性が理想的なランバート反射と異なっていても,符号の決定と領域分割の精度にほとんど影響しないことを示す.This paper describes a new method for determining the sign of Gaussian curvature at points on the object surface from three images acquired under different lighting directions. The determination can be done without knowing the directions and the strength of the light source and also surface albedo. The basic idea is well explained through the relationship between the Gaussian image of the surface and the image brightness. The proposed method doesn't need any special condition for the light source directions which was assumed in the conventional methods. It is shown that the proposed method produces good results even in the case with non-Lambertian reflectance.

  73. 投影中心に点光源がある場合のShape -from- Shading内視鏡画像からの3次元形状復元

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 1996 (5) 19-26 1996年1月18日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    投影中心に点光源がある場合に,この点光源が生成する陰影から,対象の形状を復元する手法について述べる。この条件は,内視鏡の画像から対象の形状を推定する問題と近似的に等価であることが期待される。この場合,光源と対象表面との距離が近いため,生成される陰影はこの距離に大きく依存し,問題がより困難なものになる。これを解決するため,光源からの距離が等しい表面上の曲線に注目し,この曲線の発展方程式を偏微分方程式で記述して,これを解くことで形状を復元する方法を提案する。さらに,医用内視鏡による実画像を用いて行った実験結果を示す。This paper describes an approach to reconstructing a shape from its shaded image in the case where a point light source is located at the projection center. This condition well approximates the imaging system of an endocope. In this case, the image gray level depends on not only the gradient but also the distance from the light source of each point on the surface. To deal with this difficulty, we introduce the evolution equation for equal-range contours on the surface. Propagating this contour by solving the equation, we can reconstruct a shape. Experimental results for real medical images show feasibility of this method.

  74. 固有空間法を利用した陰影からの曲面の形状復元

    岡谷 貴之, 出口 光一郎

    情報処理学会研究報告コンピュータビジョンとイメージメディア(CVIM) 1995 (33) 1-7 1995年3月23日

    出版者・発行元: 一般社団法人情報処理学会

    ISSN: 0919-6072

    詳細を見る 詳細を閉じる

    一枚の画像をもとに陰影から対象の形状を復元する手法について述べる。この問題を解くためには、対象に関する何らかの仮定が必要である。本研究では、対象の表面が局所的にあるクラスの二次曲面で近似できる、という仮定を導入する。この仮定により、局所形状の陰影だけで、凹凸の向きを除いて形状を復元できる。局所形状の復元は、二次曲面の陰影と曲面パラメータの対応をあらかじめ求めておくことで行う。この対応を表現したデータは、陰影をそのまま用いると膨大なものとなるため、陰影のKL展開に基づいた固有空間法と呼ばれる方法を用いた。This paper describes a new approach to reconstructing surface shape from its shading. By assuming that a given surface consists of a set of locally quadric segments which belong within a specific class, the surface shape is determined excepting its sign of convexity. The local segment is estimated by using the correspondence between quadric surface parameters and its shading. The correspondence is calculated prior to the estimation and stored. If this correspondence is prepared for every raw image shading, very large strage space is required. We use eigenspace method for the reduction of data size.

︎全件表示 ︎最初の5件までを表示

書籍等出版物 7

  1. 深層学習

    岡谷, 貴之

    講談社 2022年1月

    ISBN: 9784065133323

  2. 質感の科学 ―知覚・認知メカニズムと分析・表現の技術―

    小松英彦, 岡谷貴之

    朝倉書店 2016年10月20日

  3. 深層学習

    麻生 英樹, 安田 宗樹, 前田 新一, 岡野原 大輔, 岡谷 貴之, 久保 陽太郎, ボレガラ ダヌシカ

    近代科学社 2015年11月5日

  4. 深層学習(機械学習プロフェッショナルシリーズ)

    岡谷貴之

    講談社 2015年4月7日

  5. コンピュータ工学入門

    鏡 慎吾, 佐野 健太郎, 滝沢 寛之, 岡谷 貴之, 小林 広明

    コロナ社 2015年3月1日

  6. コンピュータビジョン最先端ガイド6 (CVIMチュートリアルシリーズ)

    岡谷貴之

    2013年1月1日

  7. コンピュータビジョン最先端ガイド-CVIMチュートリアルシリーズ-

    岡谷貴之, 八木康史, 斎藤英雄

    2011年4月

︎全件表示 ︎最初の5件までを表示

共同研究・競争的資金等の研究課題 25

  1. 深奥質感のマルチモーダル深層モデルの確立

    岡谷 貴之, 菅沼 雅徳, 鈴木 潤, 劉 星

    2020年11月19日 ~ 2025年3月31日

    詳細を見る 詳細を閉じる

    まず,DNNによる画像理解の研究を行い,複数の成果を得た.第一に,自然言語の指示によって作業を行うAIエージェントを開発した.エージェントが空間を見る視野をなるべく広く取るとともに,作業の指示を与える言語情報を2回解釈することに特徴があり,これによって高い作業成功率の達成が可能となった.手法は,国際的なベンチマークテストAlfredにおいて論文発表時点で全世界トップの性能を達成するとともに,成果の一端を国際会議IJACI2021にて発表した. 第二に,画像理解タスクの性能向上を目指して,画像記述の研究を行った.まず,これまでの手法が「比喩的表現」を基本的に行えないことに着目し,データセットとベンチマークテストを設計するとともに,評価方法を検討した.成果の一部は国際会議に投稿中である.第三に,画像記述のためのより高性能かつ高効率なDNNの設計を行った.画像特徴を取り出すのに,従来手法のように物体検出器に由来するCNNを使うのではなく,トランスフォーマーを用いることで高速化を達成した.同時に,物体領域から抽出した特徴と,画像を格子状に分割した領域から取り出した特徴の双方を効果的に統合することで,記述の精度を大幅に向上させている.国際的なベンチマークテストCOCO captionsにおいて,世界トップクラスの記述精度を従来手法の少なくとも数分の1の計算量で達成可能なDNNとなっている.本成果は国際会議に投稿中である.さらに,主に自然言語で表現された外部知識を,画像理解に導入する手法の検討を行っており,継続中である. また,以上とは独立に,自己教師学習の方法について研究を行った.特に近年活発に研究されている負例を用いない画像特徴の自己教師学習手法について,その有効性がどこから由来するものかを理論的に検討した.成果は国際会議に投稿中である.

  2. 深層学習がもたらすパラダイムシフトを加速するネットワーク設計理論の構築

    岡谷 貴之, 菅沼 雅徳

    2019年4月1日 ~ 2023年3月31日

    詳細を見る 詳細を閉じる

    当初計画に従い3つの項目について研究を進めた. 1つ目は画質改善タスクをターゲットとした深層ネットワークの構造の研究である.タスクを選ばないユニバーサルなネットワーク構造存在するのか,どのようにすれば実現できるかを検討したところ,一定の成果を得た.ただし訓練データと同じドメインのデータに対しては良好な性能を見せるものの,ドメインが多少なりとも変化すると性能が急激に低下する傾向があることがわかり,計画を考え直すに至った.ただしここまでの成果はarXiv上に論文として発表している. 2つ目は推論の不確かさを予測する方法である.これについては,クラス分類にコサイン類似度を用いたロジスティック回帰を行う構造をネットワークに持たせ,学習を行うことで,学習外分布(out-of-distribution)の入力を高精度に検出できることを実験的に示した.その性能は世界最高水準にあ流.以上の成果は国際会議ACCVにて発表した.この他にも,学習済みの分布と学習外分布を完全に分離して捉えるのではなく,その中間に位置するサンプルを考慮し,学習に使うことで,学習外分布の入力をより高精度に検出できることを示した.中間に位置するサンプルは,学習画像の画質を低下させる画像処理によって作られた画像を用いる.この方法は,上述の方法よりさらに高精度に,分布外入力の検出が行えることを実験を通じて示した.成果は現在国際会議に投稿中である. 3つ目は当初構想していた方法がうまく進まなかったため,方針を変更し,その検討で得た問題意識に従い,画像からひびを検出する方法の精度限界がどこにあるかを調べることを行った.訓練データ内のひびのアノテーションのエラーがひびの検出にどのように影響するかを,さまざまな角度から実験によって調べ,いくつかの知見を得た.成果は国際会議に投稿中である.

  3. 画像と言語を用いた質感情報表現のディープラーニング

    岡谷 貴之, 川嵜 圭祐, 山口 光太

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research on Innovative Areas (Research in a proposed research area)

    研究種目:Grant-in-Aid for Scientific Research on Innovative Areas (Research in a proposed research area)

    研究機関:Tohoku University

    2015年6月29日 ~ 2020年3月31日

    詳細を見る 詳細を閉じる

    画像1枚からそこに写る物の質感を認識するAIの実現という目標に一歩ずつ近づく段階的な成果を挙げた.具体的には,CNNを用いた画像ペアのランキング学習による質感認識手法,ウェブ上の画像とその説明文のペアから視覚的に認識可能な質感概念をマイニングする方法,質感認識の文脈依存性を捉えつつ商品の画像に適切なタイトルを生成する方法などである.さらに研究の過程で,質感認識は「包括的な画像認識である」と結論するに至り,今日の画像理解のためのマルチモーダル深層ニューラルネットワークの端緒となる,注意機構を用いたネットワークの開発に成功した.

  4. 機械学習技術の活用による職人的プログラミングの知能化

    滝沢 寛之, 片桐 孝洋, 横川 三津夫, 南 一生, 小林 広明, 須田 礼仁, 岡谷 貴之, 江川 隆輔, 大島 聡史

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2016年4月1日 ~ 2019年3月31日

    詳細を見る 詳細を閉じる

    本研究では,高性能計算(HPC)プログラミングの支援に機械学習を効果的に利用できる事例を示した.すでに機械学習の利用が成功している問題に変換することにより,コード最適化における種々の問題も機械学習で解決できる可能性がある.また,HPCプログラミング分野で膨大な数の訓練データを用意できる問題は稀であり,効率的な収集のためには対象問題を十分に分析する重要性が示された.さらに,HPCプログラミングと同様に,機械学習の利用においても熟練者の経験と勘に頼らなければならないが,すでに数値化されているハイパーパラメータの調整であるため,計算コストの問題に置き換えて考えることが可能であることも明らかになった.

  5. 広域災害の被害状況および復旧・復興を可視化する市街地の時空間モデリング

    岡谷 貴之

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2013年4月1日 ~ 2016年3月31日

    詳細を見る 詳細を閉じる

    車載カメラで市街地を繰り返し撮影したストリート画像系列を主な入力に,市街地の時空間モデリングを行う方法を研究し,用途や入手可能なデータの種類に応じて使い分け可能な方法を複数実現した.並行して,東日本大震災の被災地を対象に,撮影車両を走行させて映像記録を定期的に取得し,復興の過程を記録した.得られたデータは,本計画以前のものと併せて約40テラバイトである.また,同データを用いて開発した方法を検証した.以上により,同震災を対象に,震災直後の状況,月単位の復旧,年単位の復興過程それぞれの可視化と,将来起こり得る災害を対象に,被害状況を迅速に把握することの両方について,技術的な道筋を付けた.

  6. ものの質感を画像を元に認識するシステムの実現

    岡谷 貴之

    2013年4月1日 ~ 2015年3月31日

    詳細を見る 詳細を閉じる

    本年度は,質感認識の新しい手法を実現すべく,ディープラーニングを用いて質感および物体の材質カテゴリーを認識する方法を中心に,以下に述べるような研究を行った. まず,前年度構築したシステムと学習用データセットを利用して,多層畳込みニューラルネットの訓練を行い,結果を評価した.質感認識については,13の質感を表す形容詞に関する画像ペア間の順序情報を,各形容詞につき1000件程度,被験者を集めて収集し,新しいデータセットを構築した.以前われわれは,ニューラルネットを用いない方法で画像から取り出した特徴を,サポートベクトルマシンでランキングする方法を提案していた.本計画で新たに,物体認識用に訓練された畳込みニューラルネットを使って,それが画像から取り出す特徴量を代わりに用いて同様の実験を行った.その結果は,われわれが以前得ていた結果を相当上回り,ペアの順序の予測精度で10%程度向上し,トータルで90%前後の質感の認識性能を達成することが分かった. 次に,ディープ畳込みニューラルネットワークをマテリアル認識へ応用し,特に既存のデータセットFlickr Material Databse(FMD)での認識精度向上を図った.具体的には,ネットワークのモデルの検討と,転移学習やデータ拡張などの様々な手法の組み合わせを検討し,性能評価実験を繰り返し行なった.その結果,単一モデルで75%程度の認識精度を達成できた. さらに,質感認識の核でもある反射成分の物理的な分離手法の検討を行った.非負値行列分解を用いることで,従来法よりも高精度に分離が行えることを示した.同成果は国際会議ACCV2014にて発表し,Best Student Paper賞を受賞した.

  7. 画像のボケと両眼視差を同時に再現する注視反応ディスプレイ

    岡谷 貴之

    2011年4月1日 ~ 2013年3月31日

    詳細を見る 詳細を閉じる

    研究計画開始時,両眼視差とDOFボケを同時提示可能なシステムを構築し,それを用いてDOFボケが質感知覚に与える影響を調べることを研究の主要な目的としたが,初年度の23年度中には同システムの構築を終え,さらに実験もかなりの部分を終えることができたので,今年度は,さらなる詳細な検証実験の実施と,さらにこれまでの研究で見えてきた新しい課題に取り組んだ.さらに,それら成果を踏まえて,本領域の計画班や他の公募班と共同研究を実施,あるいはその準備も行った.まず取り組んだ課題は,23年度実績報告書記載の通り,質感属性の画像認識アルゴリズムの研究である.研究の結果,従来の研究に比べて,自然画像を扱えるようになり,透明感や滑らかさ,暖かさといった従来扱いにくかった質感属性を対象とできるようになった.この成果は国際会議ICPRにて公表し,さらにその内容を発展させたものを専門誌に投稿準備中である.次に,同新学術領域に参加する他の公募班(新潟大学,電気通信大学,大阪大学,京都大学の各研究グループ)と,共同研究を実施した.現在も研究中の課題を含むため詳細は省くが,本研究計画で実現した注視計測システムの生体計測への応用,われわれが実現した質感認識手法と言語的特性に基づく質感認識手法との融合による新しいアプリケーションの実現,同手法と生体の質感認識方法との比較などを含む.これらの成果は近いうちに公表される予定である.

  8. ボケを用いた立体ディスプレイシステムの研究 競争的資金

    2011年4月 ~ 2013年3月

  9. 構造化パタン光投影による植物の光合成状態計測

    岡谷 貴之

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Challenging Exploratory Research

    研究種目:Grant-in-Aid for Challenging Exploratory Research

    研究機関:Tohoku University

    2011年 ~ 2013年

    詳細を見る 詳細を閉じる

    光の反射成分を分離するコンピュータビジョンの様々な方法を,植物のクロロフィル蛍光計測の高精度化に応用する研究を行った.成果の一つは,映像プロジェクタを用いて植物にパタン光を投影し,その反射光を直接反射成分と間接反射成分に分離することで,葉の内部を経由して表面に放出される表面下散乱光成分を取り出し,クロロフィル蛍光をより高精度に計測できることを確かめたことである.この他,二色性反射モデルを使った鏡面・拡散反射成分分離の方法などを開発した.

  10. これまでの常識を上回る精度を実現する画像からの幾何学量推定方法の研究

    岡谷 貴之

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2010年 ~ 2012年

    詳細を見る 詳細を閉じる

    カメラで撮影した画像を元に,撮影対象の3次元形状をはじめとする幾何学情報を推定する問題について,その理論上最も高精度な方法を探求し,その応用を開拓した.研究成果は,行列因子分解の数値計算法,プロジェクタカメラシステムの校正法,高精度平面追跡法,車載カメラを用いた市街地の時空間モデリングと可視化の方法,車載ステレオカメラの校正法,光学的方法との融合に基づく高精度形状計測,マルコフ確率場の最適化計算法等である.

  11. 能動センシング・認識のための情報論的方法論の確立

    出口 光一郎, 岡谷 貴之, 嵯峨 聡

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2009年 ~ 2011年

    詳細を見る 詳細を閉じる

    本課題の研究では、「能動センシングのための情報学」の萌芽を結実化することが主目的で、その理論基盤が確立できた。それとともに、人間の視覚を用いたセンシング動作制御系の仕組みを明らかにし、有効情報を獲得するセンサの構造を考察して、その新しい役割を確立していく道筋を立てた。そこでの主要な導入概念は、2点間での情報の流通量を評価する相互情報量である。すなわち、運動のための視覚センサの役割を解明、動視覚センシングとも呼ぶべき新しいセンシング機構を情報の流れに注目して構築し、体系的な研究につなげていく基盤ができた。

  12. 複数プロジェクタによる高精細映像投影システムの研究 競争的資金

    2007年4月 ~ 2008年3月

  13. 複数のカメラやプロジェクタからなる映像システムのキャリブレーション手法の研究 競争的資金

    2000年4月 ~ 2008年3月

  14. 超高精細画像の投影を手軽に実現するマルチプロジェクタディスプレイの研究

    岡谷 貴之

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Young Scientists (A)

    研究種目:Grant-in-Aid for Young Scientists (A)

    研究機関:Tohoku University

    2007年 ~ 2008年

    詳細を見る 詳細を閉じる

    複数の異なるプロジェクタを組み合わせて超高解像度画像の表示を可能とするシステムを初め,プロジェクタを使ったいくつかの応用システムの研究を行った.従来は手間と時間を要していたキャリブレーション作業を画像一枚の撮影だけで済むようにしたこと,投影像を重ねて画像を高解像化する重畳投影の原理と表示映像の特性を明らかにしたこと,および映像投影によって仮想的な反射特性を立体形状の上で再現するシステムを実現したこと,などを主な成果とする.

  15. ロボットの高速で安定な行動を実現する運動視センシング機構の開発

    出口 光一郎, 岡谷 貴之, 牛田 俊, 嵯峨 智

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2006年 ~ 2008年

    詳細を見る 詳細を閉じる

    能動カメラによる追跡に基づく高速移動体の実時間追跡・形状認識システムを構築してロボットに実装し, ロボットの安定で高速な運動を実現した. また, 対象の形状や自己および対象の運動に応じたカメラの効率的な注視制御方法を考案し, 効率がどのような要因で決定されるのかなどを理論化した. 多様な状況を設定し, 多数の画像を取得して, 有効性を確認した. 以上を総合して、ロボットの高速で安定な行動を実現する運動視センシング機構の構築を完成させた.

  16. 自己校正理論に基づくプロジェクタカメラシステムの研究

    岡谷 貴之

    2005年 ~ 2006年

    詳細を見る 詳細を閉じる

    複数台の映像プロジェクタを組み合わせて大画面高精細画像を投影するシステムがある.個々のプロジェクタの解像度を台数分合計した,高解像度を実現できるため,今後の利用が期待されている.この種のシステムにおいて最も重要となるのが,最終的に生成される投影像の画質を左右する,システムのキャリブレーションである.このキャリブレーションには,2種類あり,一つが複数の異なるプロジェクタが投影する,個々の画像同士がきれいにつながるようにするための「位置合わせ」であり,もう一つが,そのようにつなぎ合わされた画像が正しいアスペクト比の正確な長方形となるようにするための「正規化」である.キャリブレーションには通常カメラを別途用い,カメラで撮影した画像が利用される.従来は,キャリブレーションの2要素のうち,「位置合わせ」についてはカメラを使った方法で実現されていたが,「正規化」は,投影先となるスクリーンの形の情報や,あるいはプロジェクタに別途取り付けた角度センサなど,カメラの画像以外の事前知識や他の情報を必要としていた・本研究では,位置合わせに加えて正規化についても,カメラで撮影した画像のみで行えることを理論的に証明するとともに,そのための具体的なアルゴリズムを導いた.これにより,プロジェクタをただ雑然と並べた後,キャリブレーション用のパタンを各プロジェクタから投影させた後,それを手持ちのカメラで一枚画像を撮影すれば,その画像のみから上述の2種類のキャリブレーションを同時に,完全に実行できるようになる.このことは,複数の映像プロジェクタによって大画面高精細画像を生成するシステムの利便性を大幅に高める.この結果,通常のフラットパネルディスプレイでは実現が難しいような,超高精細の画像を,身近なものとすることが期待される.

  17. 画像からの3次元知覚のための行動と結合した空間情報の獲得と空間再構成の研究

    出口 光一郎, 岡谷 貴之, 牛田 俊, 中島 平

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2003年 ~ 2005年

    詳細を見る 詳細を閉じる

    本研究はロボットに視覚を用いて環境を理解させる,具体的には自分の周りの3次元世界のマップを脳内に作らせるという課題に取り組んだものである.これは,実環境で動作をする実用ロボットの構成のための研究であるが,一方で,われわれ人間の空間知覚のメカニズムの解明にも挑戦しようというものであった. ここでの課題は,われわれの行動と結合した空間知覚の原理の解明とそのロボット視覚への応用である.対象の空間的な位置についての情報の幾らかは,静止している観察者によっても抽出し得る.しかし,観察者が動き始めることが,さらに多くの情報を手に入れるために必要である. 行動するということは,それがうまく実行されるために多くの空間の情報を必要とする.眼から得た画像によって矛盾のない空間情報を頭の中で再現することができれば,その空間情報のもとで行動をすると,眼から得られる画像は予想通りの変化をする.一方,この空間情報の再現が不十分であると正確な行動ができないが,それに加え,不自然な画像変化をも引き起こす.ただし,この不自然な画像変化は,それによってまた移動したその観察者に多くの情報をもたらし得る.われわれは,この不自然な変化に注目した.そしてわれわれは,ロボットのカメラ特性と行動特性をひとまとまりにしたシステムモデルをロボットに設け,この矛盾を無くすようにシステムモデルを修正することで空間を認識する方法を提案した. この原理を実ロボットに搭載したステレオ視覚システムに応用し,実環境での実験によって,本手法の有効性と将来性を確認した。

  18. 能動視覚による動的な空間知覚と立体形状認識機構の解明とその応用システムの構築

    出口 光一郎, 中島 平, 岡谷 貴之

    2003年 ~ 2003年

    詳細を見る 詳細を閉じる

    本研究では、人間の視覚の運動機構に関する知見と、研究代表者らのこれまでの能動視覚に関する研究成果を実用化に結びつけるべく、統一的な能動視覚の原理を確立するために、平成13年度に開始した特定領域研究課題を継続・発展した。その中で、本年度は特に次の3つのシステムの実現をはかった。 (1)能動カメラによる追跡に基づく高速移動体の実時間形状認識システム (2)協調する複数カメラからの画像の統合による移動体、静止物体の実時間形状認識システム (3)能動的にカメラを移動させながら撮影した多方向からの画像による高精度形状認識システム 第一のシステムでは、たとえば、移動ロボットの前方を横切る物体、監視カメラの前方を横切る車両などの立体形状を実時間で認識する。このとき、シーン中を移動する物体を発見したならば、それに視点を固定して追跡するようにカメラのパン、チルトを制御することで、立体形状が容易に計測しうる数理的な仕掛けが、平成13年度の研究よって示されている。本年度では、この原理をシステム化し実用性を確認した。 第二のシステムでは、人間の両眼を模した2台のカメラが注視点をそろえて協調して対象の動きをとらえ、形状を計測するシステムを構築した。 その結果を、多数(10台程度)のカメラによるシステムへ拡張し、いわゆる、モーションキャプチャの新しい方式を提案し、実システムとして提示する。 第三のシステムでは、自由にカメラを移動させながら対象を撮影し、得られた画像系列から対象の立体形状を再現する。画像上の対応点の動きからカメラの軌道を再現し、対象形状を復元する。対象の形状に応じたカメラの効率的な移動方法があるか、それは、どのような要因で決定されるのかなどを理論化することも、研究の目的であった。 そして、これらの成果を学術雑誌、国際会議、国内学会にて公表した。

  19. 画像からの3次元モデル取得法の研究 競争的資金

    1999年4月 ~ 2002年3月

  20. 能動視覚による動的な空間知覚と立体形状認識機構の解明とその応用システムの構築

    出口 光一郎, 中島 平, 岡谷 貴之

    2002年 ~ 2002年

    詳細を見る 詳細を閉じる

    本研究では、人間の視覚情報処理における眼球の運動制御を含めた能動的な視覚の仕掛けを解明するとともに,能動視覚をもつロボットを設計して,その視覚による空間知覚と運動制御を実現するシステムを構築する。 本年度は、このための1つのシステムの実現である,「行動と画像変化を結びつけたロボットの空間認識」について研究した. 人間は,2眼を用いることにより環境中の物体間の距離あるいは自分自身と物体までの距離を知ることができる.ただし,正確な絶対距離や,正確な方向を認識しているわけではない.対象の空間的な位置についての情報の幾らかは,静止している観察者によっても抽出し得るが正確ではない.しかし、観察者が動き始めれば,さらに多くの情報が手に入るようになる.移動を行うことは,それがうまく実行されるために多くの空間の情報を必要とする.正確な空間情報のもとで正確な行動をすると,眼から得られる画像は予想通りの変化をする.一方,この空間情報が不十分であると正確な行動ができないが,それに加え,不自然な画像変化をも引き起こす.この不自然な画像変化は,それによってまた移動したその観察者に多くの情報をもたらし得る.本研究では,このような行動と画像変化の結びつきに注目した,以下の研究を行った。 1.能動カメラによる追跡に基づく高速移動体の実時間形状認識システム 2.協調する複数カメラからの画像の統合による移動体、静止物体の実時間形状認識システム 3.ロボットに搭載した能動的にカメラによる、ロボットの空間認識 そして、これらの成果を統合し、人間の視覚情報処理における眼球の能動制御の仕掛けを解明するとともに、能動視覚をもつロボットを実現して、その視覚による空間知覚と運動制御を実現を図った。

  21. 眼底画像からの3次元眼底パターンの復元と表示の研究

    出口 光一郎, 中島 平, 本谷 秀堅, 岡谷 貴之

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    研究機関:Tohoku University

    2000年 ~ 2002年

    詳細を見る 詳細を閉じる

    本研究では眼底を多方向から撮影した複数の眼底画像を元に,眼底パターンを3次元復元・表示する.眼底とは眼球の内面のことである.眼底は人体において血管を直視できる唯一の場所のため,様々な診断をする上で非常に役立つ部位である.眼底の観察は眼のレンズを通して行なわれるため,位置表示や面積や容積などの測定が容易ではない.このため今現在,診断は見る人の主観に頼って行われている.これまでには,カメラキャリブレーションを行なった後,1組のステレオ画像から眼底を3次元表示する方法が提案されていた.しかし,1枚の眼底画像には限られた狭い範囲しか撮影できないので,この手法では3次元復元も狭い領域に限られていた.本研究では,互いに重なり合ったより多くの眼底画像から,広い範囲の3次元復元を行なう手法を提案し,実験結果を示した. 眼底画像は,カメラのレンズや患者の眼の光学系により歪んでいる.このレンズ系の特性には個人差があり校正できない.ところで眼の光学系を単レンズでモデル化し,眼底の形状がほぼ球面であることを用いると,眼底の実像は2次曲面となる.眼底画像の歪み方は,この実像に対するカメラの位置・姿勢を反映している.それぞれ歪み方の異なる眼底画像が同一の実像を撮影したものであることに基づき,各カメラの位置・姿勢および2次曲面の形状を推定し,眼底の形状を復元する.結果を3次元表示することにより,広い範囲の疾患を一度に表示できるようになり,また眼底形状の疾患の部位の特定などが容易になった. 本研究の成果が実用化されるには,まだ,医師の協力による臨床的な検証を経なければならないが,ここでの成果はコンピュータビジョン応用における新しい手法の開発とともに,新しい領域の開拓をももたらしたものである.

  22. 能動視覚による動的な空間知覚と立体形状認識機構の解明とその応用システムの構築

    出口 光一郎, 中島 平, 岡谷 貴之

    2001年 ~ 2001年

    詳細を見る 詳細を閉じる

    人間の能動的な視覚機構の数理的な構造、特に注視のための身体運動と眼球運動の視覚情報処理における役割と、立体形状知覚センサーとしての仕掛けを解明を目的として、その機構を模した自らも高速で移動したり首を振ったりできる能動的な搭載カメラによって、刻々変化する空間環境をとらえ、カメラの置かれている空間や対象の形状を実時間で認識するシステムを構築した。さらに、このとき、このような目的で認識にとってよりよい画像を得るような運動をカメラ自らが想起して行動するための、能動視覚の基本原理を確立した。 本年度は、特に次の3つのシステムの構築を目指し、一部を実現した。 (1)能動カメラによる追跡に基づく高速移動体の実時間形状認識システム-たとえば、移動ロボットの前方を横切る物体、監視カメラの前方を横切る車両などの立体形状を実時間で認識する (2)協調する複数カメラからの画像の統合による移動体、静止物体の実時間形状認識システム-人間の両眼を模した2台のカメラが注視点をそろえて協調して対象の動きをとらえ、形状を計測する (3)能動的にカメラを移動させながら撮影した多方向からの画像による高精度形状認識システム-自由にカメラを移動させながら対象を撮影し、得られた画像系列から対象の立体形状を再現する そして、これらの成果を統合し、人間の視覚情報処理における眼球の能動制御の仕掛け、能動視覚をもつロボットにおけるその視覚による空間知覚と運動制御を考察した。

  23. 画像の対応付けなどの高度な画像処理を要しない視覚サーボによるロボット制御の研究

    出口 光一郎, 岡谷 貴之, 本谷 秀堅

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B).

    研究種目:Grant-in-Aid for Scientific Research (B).

    研究機関:Tohoku University

    1998年 ~ 2000年

    詳細を見る 詳細を閉じる

    本研究の目的は、視覚センサ(カメラ)を搭載して自由に動くことが出来るロボットやマニピュレータの視覚サーボによる制御において、画像特徴を選択、抽出、そして、対応付けなどの複雑で高度な画像処理を必要としない手法を開発し、工場、現場などで実際に容易に応用できる技術として確立することであった。 本研究で開発した手法の基本原理は以下のものである。 カメラによって得られる画像を構成するデータの数(画素の数)をk、撮像条件がm個のパラメータで制御出来、そして、カメラの位置・姿勢をn個のパラメータで決めることが出来るとすると、ある条件の下でカメラがある位置である姿勢をとり、ある画像が得られるという事象はk+m+n次元の空間の一点として表される。ただし、ある対象に対しては任意の画像が得られるわけではなく、また、任意の撮像条件がとれるわけでもないので、実際に可能な実現は、この空間内のずっと次元の低い空間に制限される。このとき、わずかに条件を変えたり、カメラの位置・姿勢をわずかだけ変化させた場合、得られる画像の変化もほとんどの場合わずかである。このことは、この制限された空間が、低い次元の多様体曲面で表現されることを意味する。 カメラを動かすとどのような画像が得られるかを予測し、ある特定の画像が得られるようにするにはカメラの姿勢をどう制御すればよいか、などの基本原理をこの多様体を元に導くことが出来た。その結果、ゴール周辺の画像によって、この多様体をよく近似する平面を構成することで、制御を実現する手法を与えた。この近似平面を、与えられたサンプル画像群から主成分分析を用いて導き出す具体的な手法も開発した。それによれば、ユーザは単にこのサンプル画像群をあたえるだけ視覚サーボが実現できることも確認した。

  24. 陰と陰影を用いた画像理解に関する研究 競争的資金

    1994年4月 ~ 1999年3月

  25. 疑似ステレオ法を用いた視覚サーボによる遠隔操作の研究

    出口 光一郎, 岡谷 貴之, 本谷 秀堅

    提供機関:Japan Society for the Promotion of Science

    制度名:Grants-in-Aid for Scientific Research Grant-in-Aid for Scientific Research (B)

    研究種目:Grant-in-Aid for Scientific Research (B)

    1997年 ~ 1999年

    詳細を見る 詳細を閉じる

    本研究では、遠隔地で(あるいは時間を隔てて)ある対象の動きをとらえた動画像をもとに、同じ形状の対象に同じ動きを再現させるべく,研究代表者らが提案している「擬似ステレオ」という概念を導入して、視覚サーボによる制御を行う手法を開発した。 本研究の目的は、この原理に基づく遠隔操作のシステムを構築し、理論を検証してその有効性を確認するとともに、実験システムを構築して、実用化の課題を明らかにし、それらの解決を図るものである。単に理論研究に終らず、基礎的な理論の検証とともに、カメラと画像の伝送系も含めたロボットハンド・システムを構築し、視覚サーボによるロボットやマニピュレータの遠隔制御の実用化と応用を計ることを念頭に置いた。 特に、産業用ロボットの設計・製造に携わっている民間の専門研究者との交流による、具体的な問題の提起と、本研究の成果の応用面での検討を踏まえ、組み立て工場などの現場でも応用できるロボットの視覚による遠隔制御系のシステム開発と、その検証をはかった。 平成9年度から11年度の3年間の研究成果は、 (1)現場での応用を指向した疑似ステレオに基づく遠隔操作の実験システムを構築し、フィールドテストを模擬した研究室での実験を重ね、実験システムの有効性を検証した。 (2)ロボットの制御性能、特に、画像処理速度の制御性能への影響の検討を行い、昨今のパーソナルコンピュータ用の高速CPUを用いれば、画像処理高速化のための特別なハードウェアを必要とせずに、実用化における応用範囲をカバーできることを確認した。 (3)研究を総括し、実用的な実環境下における遠隔制御システムの構築の基礎となりうる、ロボット視覚、それに基づく行動制御の理論を確立して論文としてまとめ、学術雑誌、国際会議、国内研究会にて発表をした。

︎全件表示 ︎最初の5件までを表示

その他 2

  1. 認識の数理モデルと高階・多層確率場による高次元実データ解析

  2. 極限環境を対象とした頑健なロボット視覚の実現

    詳細を見る 詳細を閉じる

    極限環境を対象とした頑健なロボット視覚を実現する.