PoseTrack by hrnet

https://github.com/lxy5513/cvToolBox

DEMO

python pose_estimation/video.py
-o output_name
-i /path/to/video -o output_name
--display # 显示出来(display in screen)
--camera # 通过网络摄像头作为视频的输入(open web camera by video input)

ENV CONFIGURE

conda env create -f env_info_file.yml 
cd lib && make

Model Download

pose model(pose_hrnet_w48_256x192.pth) address: https://drive.google.com/drive/folders/1nzM_OBV9LbAEA7HClC0chEyf_7ECDXYA)

save in $hrnet/models/pytorch/pose_coco/pose_hrnet_w48_256x192.pth

yolov3 model download: wget https://pjreddie.com/media/files/yolov3.weights

save in $hrnet/lib/detector/yolo/yolov3.weights

UPDATE 2019-05-21

add high mAP mmdetection python pose_estimation/demo_mmd.py

do RP accuracy test cd tools && ./eval_coco.sh

---

imporove from origin code

# 通过flow net2来平滑视频(smooth pose joints by flownet2)
python pose_estimation/smooth.py

# 通过SGfilter, 最小二乘法和低次多项式平滑视频 (smooth pose joints by SG-filter)
python pose_estimation/SGfilter.py

[todo]

使用flownet2来实现视频姿态track(add tracking module by flownet2)

添加R-FCN、SSD (add other human bounding-box detector like R-FCN SSD)

---

original code clone from https://github.com/leoxiaobin/deep-high-resolution-net.pytorch