Research map / 3D vision / Human pose estimation
GRAFT: Geometric Refinement and Fitting Transformer for Human Scene Reconstruction
Pradyumna YM, Yuxuan Xue, Yue Chen, Nikita Kister, István Sárándi, Gerard Pons-Moll
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 rollout supervision (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.358 -> 0.565 | Improved |
| add geometry-grounded features (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.389 -> 0.565 | Improved |
| add explicit scale prediction (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.508 -> 0.565 | Improved |
| add GT-based training queries (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.544 -> 0.565 | Improved |
| add visual features (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.569 -> 0.565 | No clear effect |
| add visual features (ablation) | GRAFT · PROX-test (ablation, 120k steps) | Contact F1: 0.523 -> 0.572 | Improved |
| add explicit V2S loss (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.565 -> 0.597 | Improved |
| add explicit V2S loss (ablation) | GRAFT · PROX-test (ablation, 120k steps) | Contact F1: 0.572 -> 0.563 | Hurt |
| add explicit contact and penetration loss (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.565 -> 0.563 | Hurt |
| add explicit contact and penetration loss (ablation) | GRAFT · PROX-test (ablation, 120k steps) | Contact F1: 0.572 -> 0.591 | Improved |
| add iterative refinement to initialization (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | Contact F1: 0.411 -> 0.565 | Improved |
| add refinement (PA-MPJPE cost) (ablation) | GRAFT · RICH-100 (ablation, 120k steps) | PA-MPJPE: 43.51 -> 49.30 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Human3R — Human3R: Everyone Everywhere All at Once (ICLR 2026) · better on 7 of 7 shared results
- UniSH — UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass (CVPR 2026) · better on 2 of 2 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
- Reconstructing People, Places, and Cameras (CVPR 2025 · similarity 0.89)
- End-to-End HOI Reconstruction Transformer with Graph-based Encoding (CVPR 2025 · similarity 0.89)
- UniCon3R: Unified Contact-aware 4D Human-Scene Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.88)
- PICO: Reconstructing 3D People In Contact with Objects (CVPR 2025 · similarity 0.88)
- Template Free Reconstruction of Human-object Interaction with Procedural Interaction Generation (CVPR 2024 · similarity 0.88)
- Reconstructing In-the-Wild Open-Vocabulary Human-Object Interactions (CVPR 2025 · similarity 0.88)
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.
