Research map / 3D vision / Depth and stereo
Confidence Propagation through CNNs for Guided Sparse Depth Regression
Abdelrahman Eldesokey, Michael Felsberg, Fahad Shahbaz Khan
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 |
|---|---|---|---|
| enforce non-negativity of applicability via SoftPlus(beta=10) (ablation) | MS-Net[LF]-L1 (gd) · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 1042.65 -> 908.76 | Improved |
| enforce non-negativity of applicability via SoftPlus(beta=10) (ablation) | MS-Net[LF]-L1 (gd) · KITTI-Depth (selected val subset, 1k images) | MAE [mm]: 277.67 -> 209.56 | Improved |
| add output confidence to the RGB feature extraction network (ablation) | MS-Net[LF]-L2 (gd) · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 919.55 -> 870.82 | Improved |
| add depth refinement layers (ablation) | MS-Net[LF]-L2 (gd) · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 912.82 -> 870.82 | Improved |
| freeze the unguided network instead of training end-to-end (ablation) | MS-Net[LF]-L2 (gd) · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 886.09 -> 870.82 | Improved |
| late fusion instead of early fusion in multi-stream network | MS-Net · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 932.01 -> 908.76 | Improved |
| early fusion instead of late fusion in encoder-decoder network | EncDec-Net · KITTI-Depth (selected val subset, 1k images) | RMSE [mm]: 1053.91 -> 1007.71 | Improved |
| use normalized convolution with confidence propagation instead of doubled-depth standard CNN (ablation) | EncDec-Net[EF] · KITTI-Depth | RMSE [mm]: 1122.51 -> 1007.71 | Improved |
| add confidence term to the loss (versus Huber norm only) (ablation) | unguided NConv-CNN · KITTI-Depth | MAE | Improved |
| train with L2 loss instead of L1 loss | MS-Net[LF] (gd) · KITTI-Depth (test set) | RMSE [mm]: 859.22 -> 829.98 | Improved |
| train with L1 loss instead of L2 loss | MS-Net[LF] (gd) · KITTI-Depth (test set) | MAE [mm]: 233.26 -> 207.77 | Improved |
| SoftPlus instead of ReLU as non-negative function (ablation) | unguided NConv-CNN · KITTI-Depth | convergence | Improved |
Closest papers by meaning
- Uncertainty-Aware CNNs for Depth Completion: Uncertainty from Beginning to End (CVPR 2020 · similarity 0.89)
- Scale Propagation Network for Generalizable Depth Completion (TPAMI 2024 · similarity 0.88)
- Sparse Auxiliary Networks for Unified Monocular Depth Prediction and Completion (CVPR 2021 · similarity 0.88)
- VA-DepthNet: A Variational Approach to Single Image Depth Prediction (ICLR 2023 · similarity 0.87)
- Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue (ECCV 2016 · similarity 0.87)
- Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework (ICLR 2025 · similarity 0.86)
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.
