efficient deep learning for stereo matching
play

Efficient Deep Learning for Stereo Matching Wenjie Luo, Alex Schwing - PowerPoint PPT Presentation

Efficient Deep Learning for Stereo Matching Wenjie Luo, Alex Schwing and Raquel Urtasun W. Luo et.al. (UofT) Stereo Matching 1 / 6 Stereo Estimation Desired Properties: Good enough to detect obstacles precisely Fast: real time Robust to:


  1. Efficient Deep Learning for Stereo Matching Wenjie Luo, Alex Schwing and Raquel Urtasun W. Luo et.al. (UofT) Stereo Matching 1 / 6

  2. Stereo Estimation Desired Properties: Good enough to detect obstacles precisely Fast: real time Robust to: ◮ Saturation ◮ Shadows ◮ Repetitive patterns ◮ Specularities ◮ etc Can we leverage deep learning to do stereo estimation? W. Luo et.al. (UofT) Stereo Matching 2 / 6

  3. Current Deep Learning Approaches Current approaches use a siamese network Combine the two branches via concatenation follow by further processing Treat the problem as classification (i.e., given a left and right image patches, are they a true match?) [J. Zbontar and Y. LeCun, CVPR15] W. Luo et.al. (UofT) Stereo Matching 3 / 6

  4. Current Deep Learning Approaches Current approaches use a siamese network Combine the two branches via concatenation follow by further processing Treat the problem as classification (i.e., given a left and right image patches, are they a true match?) ◮ Too slow: 1 minute of computation on the GPU for KITTI! ◮ Matching not great, as scores are not correlated for different disparities [J. Zbontar and Y. LeCun, CVPR15] W. Luo et.al. (UofT) Stereo Matching 3 / 6

  5. Stereo Estimation We propose a siamese architecture with a simple product layer, which is much faster (i.e., less than 1s in GPU) Train the network with multi-class loss so that the scores are calibrated, incorporating context information p i ( y i ) ⊙ Inner product Patch representation Left image patches Right image patches (Architecture) (Learning) W. Luo et.al. (UofT) Stereo Matching 4 / 6

  6. Quantitative Results on KITTI 2015 Our approach produces much more accurate matches, 2-orders of magnitude faster than competing approaches [Zbontar & LeCun, CVPR 2015] > 2 pixel > 3 pixel > 4 pixel > 5 pixel End-Point Runtime(s) Non-Occ All Non-Occ All Non-Occ All Non-Occ All Non-Occ All MC-CNN-acrt 15.20 16.83 12.45 14.12 11.04 12.72 10.13 11.80 4.01 px 4.66 px 22.76 Ours(37) 1.84 px 2.56 px 0.34 9.96 11.67 7.23 8.97 5.89 7.62 5.04 6.78 W. Luo et.al. (UofT) Stereo Matching 5 / 6

  7. Quantitative Results on KITTI 2015 Our approach produces much more accurate matches, 2-orders of magnitude faster than competing approaches [Zbontar & LeCun, CVPR 2015] > 2 pixel > 3 pixel > 4 pixel > 5 pixel End-Point Runtime(s) Non-Occ All Non-Occ All Non-Occ All Non-Occ All Non-Occ All MC-CNN-acrt 15.20 16.83 12.45 14.12 11.04 12.72 10.13 11.80 4.01 px 4.66 px 22.76 Ours(37) 1.84 px 2.56 px 0.34 9.96 11.67 7.23 8.97 5.89 7.62 5.04 6.78 To be competitive this methods require cost-aggregation, semi-global matching follow by sophisticated post processing All/All All/Est Noc/All Noc/Est Runtime D1-bg D1-fg D1-all D1-bg D1-fg D1-all D1-bg D1-fg D1-all D1-bg D1-fg D1-all (s) MBM 4.69 13.05 6.08 4.69 13.05 6.08 4.33 12.12 5.61 4.33 12.12 5.61 0.13 SPS-St 3.84 12.67 5.31 3.84 12.67 5.31 3.50 11.61 4.84 3.50 11.61 4.84 2 MC-CNN 2.89 8.88 3.89 2.89 8.88 3.88 2.48 7.64 3.33 2.48 7.64 3.33 67 Displets v2 3.00 5.56 3.43 3.00 5.56 3.43 2.73 4.95 3.09 2.73 4.95 3.09 265 Ours(37) 3.73 8.58 4.54 3.73 8.58 4.54 3.32 7.44 4.00 3.32 7.44 4.00 1 W. Luo et.al. (UofT) Stereo Matching 5 / 6

  8. Qualitative Results on KITTI 2015 Our code is available at: http://www.cs.toronto.edu/deepLowLevelVision/ W. Luo et.al. (UofT) Stereo Matching 6 / 6

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend