Research map / Video / Video segmentation and tracking
DeepV2D: Video to Depth with Differentiable Structure from Motion
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 |
|---|---|---|---|
| Use FCRN single-image depth initialization rather than constant-depth initialization (ablation) | DeepV2D on NYUv2 · NYUv2 test | Abs Rel: 0.072 -> 0.061 | Helped |
| Replace stacked hourglass networks with single 2D and 3D hourglass networks (ablation) | DeepV2D · NYU depth evaluation | Abs-Rel: 0.065 to 0.071 | Hurt |
| Replace the 3D stereo network with correlation and a 2D encoder-decoder (ablation) | DeepV2D · NYU depth evaluation | Abs-Rel: 0.065 to 0.135 | Hurt |
| Downsample images to 1/2 resolution (ablation) | DeepV2D · NYU depth evaluation | Abs-Rel: 0.062 -> 0.083 | Hurt |
| Replace the learned motion module with identity poses (ablation) | DeepV2D · NYUv2 test | Abs-Rel: 0.062 -> 0.362 | Hurt |
| Replace the learned motion module with COLMAP poses (ablation) | DeepV2D · NYUv2 test | Abs-Rel: 0.062 -> 0.199 | Hurt |
| Replace the learned motion module with OpenMVG poses (ablation) | DeepV2D · NYUv2 test | Abs-Rel: 0.062 -> 0.173 | Hurt |
| Replace the learned motion module with PWC+Ceres poses (ablation) | DeepV2D · NYUv2 test | Abs-Rel: 0.062 -> 0.274 | Hurt |
| Use DeepV2D-predicted poses with MVSNet instead of joint depth-motion alternation (ablation) | MVSNet · NYUv2 test | Abs-Rel: 0.101 vs 0.062 | Hurt |
| Train on NYU + ScanNet rather than NYU alone (ablation) | DeepV2D on SUN3D · SUN3D test | Sc-Inv: 0.134 -> 0.104 | Helped |
Closest papers by meaning
- DeMoN: Depth and Motion Network for Learning Monocular Stereo (CVPR 2017 · similarity 0.87)
- Learning Depth From Monocular Videos Using Direct Methods (CVPR 2018 · similarity 0.87)
- VA-DepthNet: A Variational Approach to Single Image Depth Prediction (ICLR 2023 · similarity 0.86)
- DeepVideoMVS: Multi-View Stereo on Video With Recurrent Spatio-Temporal Fusion (CVPR 2021 · similarity 0.86)
- Video Depth Anything: Consistent Depth Estimation for Super-Long Videos (CVPR 2025 · similarity 0.84)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.84)
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.