Authors | Chi Su, Xiaoxuan Ma, Jiajun Su, Yizhou Wang |
Paper title | SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens |
Publication venue | arXiv preprint |
URL | https://arxiv.org/abs/2411.19824 |
Additional Information | We propose a one-stage framework for real-time multi-person 3D mesh estimation from a single RGB image. By processing input images at 644-pixel resolution and introducing scale-adaptive tokens to selectively leverage 1288-pixel features locally, our method balances high accuracy with low computational cost, enabling real-time inference. |
Full set | Kid | ||
---|---|---|---|
F1 Score | 0.950 | F1 Score | 0.730 |
Precision | 0.980 | Precision | 0.640 |
Recall | 0.910 | Recall | 0.860 |
Body | |||
---|---|---|---|
Full set | Kid | ||
MPJPE | 67.900 | MPJPE | 83.800 |
MVE | 63.300 | MVE | 79.500 |
NMJE | 71.500 | NMJE | 114.800 |
NMVE | 66.600 | NMVE | 108.900 |