envrionment configture
I use torch1.0.1 in conda
conda env create -f env_info_file.yml
DEMO
python demo.py
指定输入视频(input_video) --viz-video
指定输入视频(the keypoints of input_video) --input-npz
指定输出视频名称(the name of output video) --viz-output
指定输出的帧数(the frame number of output video) --viz-limit
RealTime Demo
https://github.com/lxy5513/videopose/blob/master/outputs/README.md
handle video by hrnet
updata at 2019-04-17
cd joints_detectors/hrnet/lib/
make
cd -
python tools/hrnet_video.py --viz-video /path/to/video.mp4
add hrnet 2D keypoints detection module, to realize the end to end 3D reconstruction
添加hrnet 2D关键点检测模块,实现更高精读的3D video 重构 hrnet
handle video by alphapose
python tools/alphapose_video.py --viz-output output.mp4 --viz-video /path/to/video.mp4
or
python tools/aphapose_video.py --viz-output output.gif --viz-video /path/to/video.mp4 --viz-limit 180
need download model
handle video with every frame keypoints
python tools/wild_video.py --viz-output output.mp4 --viz-video /path/to/video.mp4 --input-npz /path/to/input_name.npz
TODO
- add
FPN-DCNhuamn detector for hrnet to realize better accuracy.
相关模型下载reletive model download
其中hrnet依赖的pose model address: https://drive.google.com/drive/folders/1nzM_OBV9LbAEA7HClC0chEyf_7ECDXYA)
yolov3 model download: wget https://pjreddie.com/media/files/yolov3.weights
videopose model address: https://dl.fbaipublicfiles.com/video-pose-3d/cpn-pt-243.bin
paper traslation 论文翻译
https://github.com/lxy5513/videopose/blob/master/doc/translate.md
commen problems
- model download and save location
hrnet
load pose model in joints_detectors/hrnet/models/pytorch/pose_coco/ load yolov3 in joints_detectors/hrnet/lib/detector/yolo/
alphapose
load yolov3 in joints_detectors/Alphapose/models/yolo/ load duc_se joints_detectors/Alphapose/models/sppe/
videopose
for videopose model: https://github.com/lxy5513/videopose/blob/master/common/arguments.py#L29 checkpoint/