Research map / 3D vision / Depth and stereo
End-To-End Learning of Geometry and Context for Deep Stereo Regression
Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, Adam Bry
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 3-D convolutional regularization layers to unary features (ablation) | GC-Net · Scene Flow test | error >1 px (%): 97.9 -> 16.9 | Improved |
| add 3-D convolutional regularization layers to unary features (ablation) | GC-Net · Scene Flow test | MAE (px): 36.6 -> 2.51 | Improved |
| add hierarchical encoder-decoder 3-D context instead of single scale (ablation) | GC-Net · Scene Flow test | error >1 px (%): 34.6 -> 16.9 | Improved |
| add hierarchical 3-D context instead of single scale (ablation) | GC-Net · Scene Flow test | RMS (px): 20.4 -> 12.4 | Improved |
| use regression (soft argmin, L1) loss instead of hard classification loss (ablation) | GC-Net · Scene Flow test | error >1 px (%): 19.2 -> 16.9 | Improved |
| use regression loss instead of hard classification loss (ablation) | GC-Net · Scene Flow test | MAE (px): 5.01 -> 2.51 | Improved |
| use regression loss instead of soft classification loss (ablation) | GC-Net · Scene Flow test | RMS (px): 25.1 -> 12.4 | Improved |
| use classification loss instead of regression loss for training speed (ablation) | GC-Net · Scene Flow validation | training convergence | Mixed |
| use classification loss with unaries only instead of regression loss (ablation) | Unaries only · Scene Flow test | error >1 px (%): 97.9 -> 51.9 | Improved |
| use end-to-end GC-Net instead of best prior method | GC-Net · KITTI 2015 test | D1-all error (%): 3.43 -> 2.87 | Improved |
| use end-to-end GC-Net instead of Displets v2 | GC-Net · KITTI 2012 test (>3 px, all) | error rate (%): 3.09 -> 2.30 | Improved |
| use GC-Net instead of DispNetC on foreground pixels | GC-Net · KITTI 2015 test | D1-fg error (%): 4.41 -> 6.16 | Hurt |
Closest papers by meaning
- Unsupervised Adaptation for Deep Stereo (ICCV 2017 · similarity 0.90)
- Stereo Matching Using Multi-Level Cost Volume and Multi-Scale Feature Constancy (TPAMI 2019 · similarity 0.90)
- StereoDRNet: Dilated Residual StereoNet (CVPR 2019 · similarity 0.89)
- Learning Stereo from Single Images (ECCV 2020 · similarity 0.89)
- Adaptive Unimodal Cost Volume Filtering for Deep Stereo Matching (AAAI 2020 · similarity 0.89)
- A Survey on Deep Learning Techniques for Stereo-Based Depth Estimation (TPAMI 2020 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
