Research map / 3D vision / Human pose estimation
Cross-Attention of Disentangled Modalities for 3D Human Mesh Recovery with Transformers
Junhyeong Cho, Kim Youwang, Tae-Hyun Oh
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 attention masking of non-adjacent vertices (ablation) | FastMETRO-S-R50 · Human3.6M | MPJPE: 58.0 -> 55.7 | Improved |
| use graph convolutions instead of attention masking (ablation) | FastMETRO-S-R50 · Human3.6M | MPJPE: 55.7 -> 56.6 | No clear effect |
| mask attention in only half of the heads (ablation) | FastMETRO-S-R50 · Human3.6M | MPJPE: 55.7 -> 55.8 | No clear effect |
| use precomputed upsampling matrix instead of learnable linear layers (ablation) | FastMETRO-S-R50 · Human3.6M | MPJPE: 58.1 -> 55.7 | Improved |
| add progressive dimensionality reduction (ablation) | FastMETRO-S-R50 · Human3.6M | params with similar accuracy: 39.5M -> 32.7M | Improved |
| use fixed instead of learnable positional encodings (ablation) | FastMETRO-S-R50 · Human3.6M | PA-MPJPE: 39.7 -> 39.4 | Improved |
| add camera token in the transformer decoder (ablation) | FastMETRO-S-R50 · Human3.6M | PA-MPJPE: 39.5 -> 39.4 | No clear effect |
| predict camera from estimated mesh (ablation) | FastMETRO-S-R50 · Human3.6M | MPJPE: 55.7 -> 58.0 | Hurt |
| full FastMETRO design vs naive encoder-decoder baseline (ablation) | transformer encoder-decoder · Human3.6M | PA-MPJPE: 41.8 -> 39.4 | Improved |
| scale transformer from S to L variant | FastMETRO R50 · Human3.6M | PA-MPJPE: 39.4 -> 37.3 | Improved |
| swap ResNet-50 for HRNet-W64 backbone | FastMETRO-L · Human3.6M | PA-MPJPE: 37.3 -> 33.7 | Improved |
| train longer, 60 to 200 epochs | METRO-H64 · Human3.6M | PA-MPJPE: 38.0 -> 36.7 | Improved |
Closest papers by meaning
- Deformable Mesh Transformer for 3D Human Mesh Recovery (CVPR 2023 · similarity 0.89)
- One-Stage 3D Whole-Body Mesh Recovery With Component Aware Transformer (CVPR 2023 · similarity 0.88)
- Human Mesh Recovery From Monocular Images via a Skeleton-Disentangled Representation (ICCV 2019 · similarity 0.88)
- MEGA: Masked Generative Autoencoder for Human Mesh Recovery (CVPR 2025 · similarity 0.87)
- SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens (CVPR 2025 · similarity 0.86)
- GenHMR: Generative Human Mesh Recovery (AAAI 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.
