This is the official repo of CrowdPose: Efficient Crowded Scenes Pose Estimation and A New Benchmark. Our proposed method surpasses the state-of-the-art methods on CrowdPose dataset by 5 mAP and results on MSCOCO dataset demonstrate the generalization ability of our method (comparatively 0.8 mAP higher). Images in our proposed CrowdPose dataset have a uniform distribution of Crowd Index among [0, 1].
We provide evaluation tools for CrowdPose dataset. Our evaluation tools is developed based on @cocodataset/cocoapi. The source code of our model will be integrated into AlphaPose soon.
Train + Validation + Test Images (Google Drive)
Annotations (Google Drive)
Results on CrowdPose Validation:
Compare with state-of-the-art methods
| Method | AP @0.5:0.95 | AP @0.5 | AP @0.75 | AR @0.5:0.95 | AR @0.5 | AR @0.75 |
|---|---|---|---|---|---|---|
| Detectron (Mask R-CNN) | 57.2 | 83.5 | 60.3 | 65.9 | 89.3 | 69.4 |
| Simple Pose (Xiao et al.) | 60.8 | 81.4 | 65.7 | 67.3 | 86.3 | 71.8 |
| Ours | 66.0 | 84.2 | 71.5 | 72.7 | 89.5 | 77.5 |
Compare with open-source systems
| Method | AP @Easy | AP @Medium | AP @Hard | FPS |
|---|---|---|---|---|
| OpenPose (CMU-Pose) | 62.7 | 48.7 | 32.3 | 5.3 |
| Detectron (Mask R-CNN) | 69.4 | 57.9 | 45.8 | 2.9 |
| Ours | 75.5 | 66.3 | 57.4 | 10.1 |
Results on MSCOCO Validation:
| Method | AP @0.5:0.95 | AR @0.5:0.95 |
|---|---|---|
| Detectron (Mask R-CNN) | 64.8 | 71.1 |
| Simple Pose (Xiao et al.) | 69.8 | 74.1 |
| AlphaPose | 70.9 | 76.4 |