zongmianli / Realtime_Multi-Person_Pose_Estimation

Code repo for realtime multi-person pose estimation in CVPR'17 (Oral)

Realtime Multi-Person Pose Estimation

By Zhe Cao, Tomas Simon, Shih-En Wei, Yaser Sheikh.

This is a fork of the original CVPR'17 implementation of Openpose (without hand and facial landmarks) adapted to the 3D motion-force estimator.

This version is different from the official releases in the following aspects:

• A different I/O interface:
  • The input is required to be an image folder (explained here) containing a mixture of still images and video frame images.
  • The estimated 2D poses are output in the data structure required by the contact recognizer and the 3D motion-force estimator.
• Adaptation to the scenario of object manipulation in instructional videos. The following post-processing steps are appended to the testing module of Openpose:
  • We assume that there is at most one person in each input image / video frame. When multiple human instances are present, only the one detected with the highest score are preserved — the others are ignored.
  • Due to the heavy occlusion during person-object interaction, the predicted Fart Affinity Fields (PAFs) may not be correct. As a result, some joints (often hands, ankles) may be missing or mis-detected, e.g. associated to another people in the background. To address this problem we have made changes in the bottom-up parsing step in the original implementation.

Installation

git clone https://github.com/zongmianli/Realtime_Multi-Person_Pose_Estimation ~/Openpose-video

See the original repo for more information if needed.

Testing

Assume that we have a mixture of still images and videos to test.

1. Follow Step 1-3 in this doc to create a new image folder with a arbitrary name, for example, sample_imagefolder, and put the folder to the path ~/Openpose-video/testing/sample_imagefolder.

2. Go to ~/Openpose-video/testing/python/ and run:

   img_dir=~/Openpose-video/testing/sample_imagefolder
   vis_dir=~/Openpose-video/testing/sample_imagefolder_vis
   save_path=${vis_dir}/Openpose-video.pkl
   python run_imagefolder.py ${img_dir} ${vis_dir} ${save_path} --save-after-each-iteration
   

The estimated 2d poses and the corresponding visualization images will be saved in ${save_path} and ${vis_dir}, respectively.