Model | Backbone | Lr schd | mAP | AP50 | AP75 | APM | APL | Config | Download |
---|---|---|---|---|---|---|---|---|---|
PETR | R-50 | 100e | 68.8 | 87.5 | 76.3 | 62.7 | 77.7 | config | Google Drive | BaiduYun |
PETR | R-101 | 100e | 70.0 | 88.5 | 77.5 | 63.6 | 79.4 | config | Google Drive | BaiduYun |
PETR | Swin-L | 100e | 73.1 | 90.7 | 80.9 | 67.2 | 81.7 | config | Google Drive | BaiduYun |
Model | Backbone | Lr schd | Flip test | mAP | AP50 | AP75 | APE | APM | APH | Config | Download |
---|---|---|---|---|---|---|---|---|---|---|---|
PETR | Swin-L | 100e | N | 71.7 | 90.0 | 78.3 | 77.5 | 72.0 | 65.8 | config | Google Drive | BaiduYun |
PETR | Swin-L | 100e | Y | 72.3 | 90.8 | 78.8 | 78.7 | 72.9 | 65.5 | config | Google Drive | BaiduYun |
- Swin-L are trained with batch size 16 due to GPU memory limitation.
- The performance is unstable.
PETR
may fluctuate about 0.2 mAP.
@inproceedings{shi2022end,
title={End-to-End Multi-Person Pose Estimation With Transformers},
author={Shi, Dahu and Wei, Xing and Li, Liangqi and Ren, Ye and Tan, Wenming},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={11069--11078},
year={2022}
}