Research map / Video / Video generation
Diffusion-based 3D Hand Motion Recovery with Intuitive Physics
Yufei Zhang, Zijun Cui, Jeffrey O. Kephart, Qiang Ji
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 |
|---|---|---|---|
| replace deterministic Transformer refiner with conditional diffusion refiner (ablation) | K-Hand + motion refinement · DexYCB | MJE / ACCL: MJE 22.7 -> 21.9 mm; ACCL 2.53 -> 1.38 | Improved |
| add motion-state prediction loss (four hand motion states) (ablation) | diffusion refiner on K-Hand · DexYCB | MJE / ACCL: MJE 21.9 -> 21.6 mm; ACCL 1.38 -> 1.27 | Improved |
| add least-kinetic-motion loss (ablation) | diffusion refiner with state loss · DexYCB | ACCL / KIN: ACCL 1.27 -> 1.17; KIN 0.06 -> 0.00 | Improved |
| add least-kinetic-motion loss (ablation) | diffusion refiner with state loss · DexYCB | MJE: 21.6 -> 21.7 | No clear effect |
| add stability loss instead of kinetic loss (ablation) | diffusion refiner with state loss · DexYCB | MJE / KIN / STA: MJE 21.6 -> 21.5; KIN 0.06 -> 0.05; STA 0.00 | Improved |
| apply Gaussian SmoothFilter to frame-wise estimates (ablation) | K-Hand · DexYCB | MJE / ACCL: MJE 24.4 -> 25.3; ACCL 14.41 -> 1.96 | Mixed |
| use full physics losses instead of constant acceleration loss (ablation) | diffusion refiner on K-Hand · DexYCB | MJE / ACCL: MJE 21.7 -> 21.5; ACCL 1.29 -> 1.17 | Improved |
| apply physics diffusion refiner on top of HaMer instead of PoseBERT | HaMer · DexYCB | MJE / ACCL: MJE 18.0 -> 17.5 mm; ACCL 2.38 -> 1.01 | Improved |
| train refiner on HaMer's own predictions (model-specific) | physics diffusion refiner · DexYCB | MJE / ACCL: MJE 17.5 -> 12.4 mm; ACCL 1.01 -> 0.80 | Improved |
| add refiner trained only on DexYCB, tested zero-shot | HaMer · HO3Dv2 | P-MJE / F@5: P-MJE 8.1 -> 8.0; F@5 58.0 -> 59.7 | Improved |
| add physics diffusion refiner without fine-tuning | HaMer · TACO S1 (bimanual) | P-MJE / ACCL: P-MJE 9.7 -> 8.9; ACCL 9.88 -> 0.81 | Improved |
| use physics diffusion refiner instead of PoseBERT | HaMer · TACO S1 (bimanual) | P-MJE / ACCL: P-MJE 11.0 -> 8.9; ACCL 3.58 -> 0.81 | Improved |
Closest papers by meaning
- The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction (NeurIPS 2026 · similarity 0.90)
- Dyn-HaMR: Recovering 4D Interacting Hand Motion from a Dynamic Camera (CVPR 2025 · similarity 0.88)
- HOLD: Category-agnostic 3D Reconstruction of Interacting Hands and Objects from Video (CVPR 2024 · similarity 0.88)
- Physics-Aware Hand-Object Interaction Denoising (CVPR 2024 · similarity 0.88)
- Re-HOLD: Video Hand Object Interaction Reenactment via adaptive Layout-instructed Diffusion Model (CVPR 2025 · similarity 0.88)
- ComPose: When to Trust Hands for Object Pose Tracking (NeurIPS 2026 · 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.
