News!

• Aug 2020: v0.4.0 version of AlphaPose is released! Stronger tracking! Include whole body(face,hand,foot) keypoints! Colab now available.
• Dec 2019: v0.3.0 version of AlphaPose is released! Smaller model, higher accuracy!
• Apr 2019: MXNet version of AlphaPose is released! It runs at 23 fps on COCO validation set.
• Feb 2019: CrowdPose is integrated into AlphaPose Now!
• Dec 2018: General version of PoseFlow is released! 3X Faster and support pose tracking results visualization!
• Sep 2018: v0.2.0 version of AlphaPose is released! It runs at 20 fps on COCO validation set (4.6 people per image on average) and achieves 71 mAP!

AlphaPose

AlphaPose is an accurate multi-person pose estimator, which is the first open-source system that achieves 70+ mAP (75 mAP) on COCO dataset and 80+ mAP (82.1 mAP) on MPII dataset. To match poses that correspond to the same person across frames, we also provide an efficient online pose tracker called Pose Flow. It is the first open-source online pose tracker that achieves both 60+ mAP (66.5 mAP) and 50+ MOTA (58.3 MOTA) on PoseTrack Challenge dataset.

AlphaPose supports both Linux and Windows!

COCO 17 keypoints

Halpe 26 keypoints + tracking

Halpe 136 keypoints + tracking

Coco_wholebody 133 keypoints

Results

Pose Estimation

Results on COCO test-dev 2015:

Method	AP @0.5:0.95	AP @0.5	AP @0.75	AP medium	AP large
OpenPose (CMU-Pose)	61.8	84.9	67.5	57.1	68.2
Detectron (Mask R-CNN)	67.0	88.0	73.1	62.2	75.6
AlphaPose	73.3	89.2	79.1	69.0	78.6

Results on MPII full test set:

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Ave
OpenPose (CMU-Pose)	91.2	87.6	77.7	66.8	75.4	68.9	61.7	75.6
Newell & Deng	92.1	89.3	78.9	69.8	76.2	71.6	64.7	77.5
AlphaPose	91.3	90.5	84.0	76.4	80.3	79.9	72.4	82.1

Results on CocoWholebody validation set:

gt mAP

Method	Body	Foot	Face	Hand	Fullbody
AlphaPose	52.5	36.8	69.1	3.7	32.1

rcnn mAP

Method	Body	Foot	Face	Hand	Fullbody
AlphaPose	51.5	36.8	67.8	3.7	31.1

More results and models are available in the docs/MODEL_ZOO.md.

Pose Tracking

Please read trackers/README.md for details.

CrowdPose

Please read docs/CrowdPose.md for details.

Installation

Please check out docs/INSTALL.md

Model Zoo

Please check out docs/MODEL_ZOO.md

Quick Start

• Colab: We provide a colab example for your quick start.

• Inference: Inference demo

./scripts/inference.sh ${CONFIG} ${CHECKPOINT} ${VIDEO_NAME} # ${OUTPUT_DIR}, optional

• Training: Train from scratch

./scripts/train.sh ${CONFIG} ${EXP_ID}

• Validation: Validate your model on MSCOCO val2017

./scripts/validate.sh ${CONFIG} ${CHECKPOINT}

Examples:

Demo using FastPose model.

./scripts/inference.sh configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml pretrained_models/fast_res50_256x192.pth ${VIDEO_NAME}
#or
python scripts/demo_inference.py --cfg configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint pretrained_models/fast_res50_256x192.pth --indir examples/demo/

Train FastPose on mscoco dataset.

./scripts/train.sh ./configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml exp_fastpose

More detailed inference options and examples, please refer to GETTING_STARTED.md

Coco_wholebody Dataset

Downloading

Images can be downloaded from COCO 2017 website

COCO-WholeBody annotations for Train / Validation (Google Drive).

Training from scratch

• Change the dataset path in this file.
• Run the following script :

cd scripts 
python3 train.py --cfg configs/coco_wholebody/resnet/256x192_res50_lr1e-3_1x.yaml --exp-id 2

Validation

Validate your model on coco_wholebody dataset Run the following script:

 cd scripts
 python3 validate.py --cfg configs/coco_wholebody/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint {CHECKPOINT-PATH} --batch 64

Inference

• For a single image:

cd scripts
python3 demo_inference.py --cfg configs/coco_wholebody/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint {CHECKPOINT-PATH} --image {IMAGE-PATH}  

• For image directory:

cd scripts
python3 demo_inference.py --cfg configs/coco_wholebody/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint {CHECKPOINT-PATH} --indir {IMAGE_DIR-PATH}  

• For a video:

cd scripts
python3 demo_inference.py --cfg configs/coco_wholebody/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint {CHECKPOINT-PATH} --video {VIDEO-PATH} 

Common issue & FAQ

Check out faq.md for faq. If it can not solve your problems or if you find any bugs, don't hesitate to comment on GitHub or make a pull request!

Contributors

AlphaPose is based on RMPE(ICCV'17), authored by Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai and Cewu Lu, Cewu Lu is the corresponding author. Currently, it is maintained by Jiefeng Li*, Hao-shu Fang*, Yuliang Xiu and Chao Xu.

The main contributors are listed in doc/contributors.md.

TODO

• Multi-GPU/CPU inference
• 3D pose
• add tracking flag
• PyTorch C++ version
• Add MPII and AIC data
• dense support
• small box easy filter
• Crowdpose support
• Speed up PoseFlow
• Add stronger/light detectors and the mobile pose
• High level API

We would really appreciate if you can offer any help and be the contributor of AlphaPose.

Citation

Please cite these papers in your publications if it helps your research:

@inproceedings{fang2017rmpe,
  title={{RMPE}: Regional Multi-person Pose Estimation},
  author={Fang, Hao-Shu and Xie, Shuqin and Tai, Yu-Wing and Lu, Cewu},
  booktitle={ICCV},
  year={2017}
}

@article{li2018crowdpose,
  title={CrowdPose: Efficient Crowded Scenes Pose Estimation and A New Benchmark},
  author={Li, Jiefeng and Wang, Can and Zhu, Hao and Mao, Yihuan and Fang, Hao-Shu and Lu, Cewu},
  journal={arXiv preprint arXiv:1812.00324},
  year={2018}
}

@inproceedings{xiu2018poseflow,
  author = {Xiu, Yuliang and Li, Jiefeng and Wang, Haoyu and Fang, Yinghong and Lu, Cewu},
  title = {{Pose Flow}: Efficient Online Pose Tracking},
  booktitle={BMVC},
  year = {2018}
}

@inproceedings{jin2020whole,
title={Whole-Body Human Pose Estimation in the Wild},
author={Jin, Sheng and Xu, Lumin and Xu, Jin and Wang, Can and Liu, Wentao and Qian, Chen and Ouyang, Wanli and Luo, Ping},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},    
year={2020}
}

License

AlphaPose is freely available for free non-commercial use, and may be redistributed under these conditions. For commercial queries, please drop an e-mail at mvig.alphapose[at]gmail[dot]com and cc lucewu[[at]sjtu[dot]edu[dot]cn. We will send the detail agreement to you.