Research map / 3D vision / Depth and stereo
VA-DepthNet: A Variational Approach to Single Image Depth Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 V-layer (first-order variational constraint) in place of a single conv layer (ablation) | Swin-L encoder-decoder depth network · NYU Depth V2 test | SILog: 8.830 -> 8.198 with refinement (conv w/ vs. V-layer w/) | Helped |
| replace V-layer with self-attention layer (ablation) | Swin-L depth network · NYU Depth V2 test | SILog: self-attention w/ refine 8.595 vs V-layer w/ refine 8.198 | Helped |
| add confidence weight matrix Sigma to V-layer (ablation) | V-layer · NYU Depth V2 test | SILog: 8.537 -> 8.198 | Helped |
| use fixed difference operator P instead of learnable P (ablation) | V-layer · NYU Depth V2 test | SILog: 8.355 -> 8.198 | Helped |
| increase number of feature groups S from 1 to 16 (ablation) | V-layer · NYU Depth V2 test | SILog: 8.456 -> 8.198 | Helped |
| increase S from 16 to 128 (ablation) | V-layer · NYU Depth V2 test | SILog: 8.198 -> 8.172 at 2.5x lower FPS | No clear effect |
| apply V-layer at stride 8 instead of stride 16 (ablation) | V-layer · NYU Depth V2 test | SILog: 8.198 -> 8.149 with FPS dropping from 7.032 to 2.597 | No clear effect |
| add V-layer to backbones (Swin-S, ConvNeXt-S, Swin-L) (ablation) | encoder-decoder depth networks · NYU Depth V2 test | SILog: Improvement largest on weaker backbones (values only in figure) | Helped |
| add V-layer in place of convolution (ablation) | Swin-L depth network · KITTI Eigen test | SILog: 6.996 -> 6.817 | Helped |
| add V-layer in place of convolution (ablation) | Swin-L depth network · SUN RGB-D test | SILog: 13.172 -> 12.596 | Helped |
| V-layer with Swin-L (full method) vs prior SOTA NeWCRFs | Swin-L depth network · NYU Depth V2 test | SILog: 9.102 -> 8.198 | Helped |
| V-layer with Swin-L vs NeWCRFs on KITTI Eigen | Swin-L depth network · KITTI Eigen test | SILog: 6.986 -> 6.817 | Helped |
Closest papers by meaning
- Single Image Depth Prediction Made Better: A Multivariate Gaussian Take (CVPR 2023 · similarity 0.90)
- Depth from a Single Image by Harmonizing Overcomplete Local Network Predictions (NeurIPS 2016 · similarity 0.87)
- SDC-Depth: Semantic Divide-and-Conquer Network for Monocular Depth Estimation (CVPR 2020 · similarity 0.87)
- Deep Ordinal Regression Network for Monocular Depth Estimation (CVPR 2018 · similarity 0.87)
- Bilateral Cyclic Constraint and Adaptive Regularization for Unsupervised Monocular Depth Prediction (CVPR 2019 · similarity 0.87)
- Sparse Auxiliary Networks for Unified Monocular Depth Prediction and Completion (CVPR 2021 · similarity 0.87)
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.