Research map / 3D vision / Human pose estimation
PosePlaner: Denoising Any Feedforward Pose Predictor from Pairwise Planar Geometry
Yihang Chen, Bin Tan, Yujun Shen, Yinghao Xu, Nan Xue
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PosePlaner, a method that denoises feedforward multi-view pose predictors (e.g., DUSt3R, VGGT) using pairwise planar geometry, based only on the abstract provided.
Problem
Multi-view feedforward Transformers such as DUSt3R and VGGT subsume the classical SfM/SLAM pipeline in one pass, which hides the interfaces between correspondences, geometry, and pose, making their pose estimates entangled and hard to refine.
Approach
Exposes the intermediate pipeline interfaces by denoising feedforward pose predictions using pairwise planar geometry, applicable to any feedforward pose predictor. Details beyond the abstract are unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| refine weak LoFTR poses with PosePlaner | LoFTR · ScanNet1500 | rotation AUC@5: 11.5 -> 29.4 | Improved |
| refine strong ROMA/MASt3R poses with PosePlaner | ROMA, MASt3R · ScanNet1500 | rotation AUC@5: 66.7 -> 64.1 (ROMA); 66.5 -> 65.0 (MASt3R) | Hurt |
| refine streaming estimators at long sequence length | CUT3R · ScanNet streaming N=2000 | rotation AUC@5: 17.3 -> 37.3 | Improved |
| refine streaming estimators at short sequence length | TTT3R · ScanNet streaming N=200 | translation AUC@5: 11.1 -> 10.8 | No clear effect |
| replace NopeSAC head with PosePlaner | correspondence-based solvers · ScanNet1500 | rotation AUC@5: 34.9 -> 48.9 (SuperGlue) | Improved |
| use PosePlaner instead of bundle adjustment | LoFTR init · ScanNet1500 | rotation AUC@5: 13.4 -> 29.5 | Improved |
| use PosePlaner instead of bundle adjustment | MASt3R init · ScanNet1500 | rotation AUC@5: 66.5 -> 64.7 | Hurt |
| add geometry tokens g_m as conditioning (ablation) | PosePlaner · ScanNet1500 (LoFTR init) | rotation AUC@5: 11.5 -> 29.4 | Improved |
| add pose token c (upstream init) as conditioning (ablation) | PosePlaner · ScanNet1500 | rotation AUC@5: 17.4 -> 48.9 (SuperGlue init) | Improved |
| add score-head fusion instead of uniform averaging (ablation) | PosePlaner · ScanNet1500 (LoFTR init) | rotation AUC@5: 26.5 -> 29.4 | Improved |
| add translation channel to the diffusion residual (ablation) | PosePlaner · ScanNet1500 (LoFTR init) | rotation AUC@5: 16.4 -> 29.4 | Improved |
| use 6D rotation instead of unit quaternion (ablation) | PosePlaner · ScanNet1500 (LoFTR init) | rotation AUC@5: 20.4 -> 29.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 5 of 5 on most shared results.
- WinT3R — WinT3R: Window-Based Streaming Reconstruction with Camera Token Pool (ICLR 2026) · better on 1 of 1 shared result
- TTT3R — TTT3R: 3D Reconstruction as Test-Time Training (ICLR 2026) · better on 1 of 1 shared result
- DiffusionSfM — DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion (CVPR 2025) · better on 1 of 1 shared result
- PoseDiffusion — PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment (ICCV 2023) · better on 1 of 1 shared result
- SuperGlue — SuperGlue: Learning Feature Matching With Graph Neural Networks (CVPR 2020) · better on 1 of 1 shared result
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.
Experimental setup
Builds on: DUSt3R VGGT
Closest papers by meaning
- PlaneRecTR++: Unified Query Learning for Joint 3D Planar Reconstruction and Pose Estimation (TPAMI 2025 · similarity 0.90)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.88)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.88)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.87)
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 2026 · similarity 0.87)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 2025 · 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.
