Research map / 3D vision / Depth and stereo
$\pi^3$: Permutation-Equivariant Visual Geometry Learning
Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, Tong He
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 affine-invariant camera pose modeling (removes reference view, permutation-equivariant) (ablation) | Pi3 ablation · ETH3D point map | Accuracy mean (lower better): 0.197 -> 0.131 | Improved |
| add scale-invariant local pointmaps (ablation) | Pi3 ablation · ETH3D point map | Accuracy mean: 0.229 -> 0.197 | Improved |
| add scale-invariant local pointmaps (ablation) | Pi3 ablation · 7-Scenes / NRGBD indoor | Accuracy: 7-Scenes 0.020 -> 0.020 | No clear effect |
| add affine-invariant camera pose modeling (ablation) | Pi3 ablation · 7-Scenes | Accuracy mean: 0.020 -> 0.019 | No clear effect |
| replace reference-view design with permutation-equivariant design | VGGT-style baseline · DTU input-order robustness | Acc. std: 0.033 -> 0.003 | Improved |
| replace reference-view design with permutation-equivariant design | VGGT-style baseline · ETH3D input-order robustness | Acc. std: 0.049 -> 0.000 | Improved |
| train from scratch with only core objectives (no global proxy) | Pi3 · ETH3D point map | Accuracy: 0.563 (VGGT) vs 0.618 | Hurt |
| add auxiliary global pointmap proxy head during from-scratch training | Pi3 · ETH3D point map | Accuracy: 0.618 -> 0.418 | Improved |
| add auxiliary global proxy head | Pi3 · 7-Scenes | Completion: 0.068 -> 0.071 | No clear effect |
| Pi3 vs VGGT on camera pose | Pi3 · Sintel | ATE: 0.167 -> 0.074 | Improved |
| Pi3 vs VGGT on video depth | Pi3 · Sintel | Abs Rel: 0.299 -> 0.233 | Improved |
| Pi3 vs VGGT on TUM-dynamics pose | Pi3 · TUM-dynamics | ATE: 0.012 vs 0.014 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 12 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Neural Isometries: Taming Transformations for Equivariant ML (NeurIPS 2024 · similarity 0.87)
- Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation (NeurIPS 2024 · similarity 0.87)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.86)
- EquiPose: Exploiting Permutation Equivariance for Relative Camera Pose Estimation (CVPR 2025 · similarity 0.86)
- 3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction (NeurIPS 2024 · similarity 0.86)
- VGGT: Visual Geometry Grounded Transformer (CVPR 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.
