Research map / 3D vision / Depth and stereo
Depth Estimation via Affinity Learned with Convolutional Spatial Propagation Network
Xinjing Cheng, Peng Wang, Ruigang Yang
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 mirror connections (UNet) to the depth network, concatenating after relu | Ma et al. network · NYU v2 (color only) | RMSE: 0.230 -> 0.137 | Improved |
| replace SPN with CSPN for refinement | Ma et al. network · NYU v2 (color only) | RMSE: 0.172 -> 0.162 | Improved |
| replace SPN with CSPN with sparse depth embedded | UNet · NYU v2 (500 sparse samples, preserve SD) | RMSE: 0.144 -> 0.117 | Improved |
| add bilateral filtering post-processing | Ma et al. network · NYU v2 (color only) | RMSE: 0.230 -> 0.479 | Hurt |
| add UNet mirror-connection network to SPN with sparse depth | SPN with sparse depth · NYU v2 (500 sparse samples, preserve SD) | RMSE: 0.144 vs 0.137 | No clear effect |
| add CSPN with sparse depth to the UNet network | UNet · KITTI (500 sparse samples) | delta<1.02 (%): 62.6 -> 70.2 | Improved |
| add SPN with sparse depth to the UNet network | UNet · KITTI (500 sparse samples) | RMSE: 3.049 -> 3.248 | Hurt |
| concatenate mirror connection after relu instead of after conv (ablation) | UNet mirror connection · NYU v2 (color only) | RMSE: 0.531 -> 0.137 | Improved |
| concatenate mirror connection after relu instead of after bn (ablation) | UNet mirror connection · NYU v2 (color only) | RMSE: 0.158 -> 0.137 | Improved |
| increase CSPN iterations up to 24 (kernel 3) (ablation) | CSPN · NYU v2 | RMSE: accuracy is saturated when the number of iterations is increased to 24 | Improved |
| increase CSPN kernel size beyond 7 to 9 (N=12) (ablation) | CSPN · NYU v2 | RMSE: saturated at size of 7; performance drops slightly at 9 | Hurt |
| use CSPN instead of SPN for propagation speed | refinement module · 1024x768 image, 4 iterations | ms: 127.902 -> 3.689 | Improved |
Closest papers by meaning
- Learning Depth with Convolutional Spatial Propagation Network (TPAMI 2019 · similarity 0.94)
- CSPN++: Learning Context and Resource Aware Convolutional Spatial Propagation Networks for Depth Completion (AAAI 2020 · similarity 0.91)
- Dynamic Spatial Propagation Network for Depth Completion (AAAI 2022 · similarity 0.90)
- Non-Local Spatial Propagation Network for Depth Completion (ECCV 2020 · similarity 0.89)
- Bilateral Propagation Network for Depth Completion (CVPR 2024 · similarity 0.89)
- Monocular Depth Estimation with Affinity, Vertical Pooling, and Label Enhancement (ECCV 2018 · similarity 0.88)
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.
