Research map / 3D vision / Depth and stereo
Input-Level Inductive Biases for 3D Reconstruction
Wang Yifan, Carl Doersch, Relja Arandjelović, João Carreira, Andrew Zisserman
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 camera geometry embeddings (ablation) | Perceiver IO depth estimation · ScanNet | absolute relative difference: 0.2568 -> 0.1350 (30%); 0.2423 -> 0.1293 (50%); 0.2340 -> 0.1234 (100%) | Improved |
| Add epipolar embeddings without camera embeddings (ablation) | Perceiver IO depth estimation · ScanNet | absolute relative difference: 0.2568 -> 0.2084 (30%); 0.2423 -> 0.2018 (50%); 0.2340 -> 0.1853 (100%) | Improved |
| Add epipolar embeddings alongside camera embeddings (ablation) | Perceiver IO depth estimation · ScanNet | absolute relative difference: 0.1350 -> 0.1371 (30%); 0.1293 -> 0.1304 (50%); 0.1234 -> 0.1204 (100%) | Mixed |
| Use camera center and ray direction rather than projection matrix and pixel coordinates (ablation) | IIB · ScanNet | absolute relative difference: 0.1345 -> 0.1234 (channels); 0.1805 -> 0.1249 (tokens) | Improved |
| Concatenate camera geometry as channels rather than separate tokens (ablation) | IIB · ScanNet | absolute relative difference: 0.1249 -> 0.1234; 0.1805 -> 0.1345 | Improved |
| Use relative-angle rather than normal-vector epipolar encoding (ablation) | IIB · ScanNet | absolute relative difference: 0.1235 -> 0.1204 | Improved |
| Remove RGB information from output queries (ablation) | IIB · ScanNet | validation error | Improved |
| Train on ScanNet and evaluate on similar-domain SUN3D instead of training on SUN3D | IIB · SUN3D | absolute relative difference: 0.0985 -> 0.1291 | Hurt |
| Train on ScanNet and evaluate on RGBD-SLAM instead of training on RGBD-SLAM | IIB · RGBD-SLAM | absolute relative difference: 0.0951 -> 0.2572 | Hurt |
Closest papers by meaning
- $SE(3)$ Equivariant Ray Embeddings for Implicit Multi-View Depth Estimation (NeurIPS 2024 · similarity 0.87)
- The More You See in 2D the More You Perceive in 3D (CVPR 2024 · similarity 0.87)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.86)
- Déjà View: Looping Transformers for Multi-View 3D Reconstruction (NeurIPS 2026 · similarity 0.86)
- Cameras as Relative Positional Encoding (NeurIPS 2025 · similarity 0.86)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 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.
