search

Lotayou / everybody_dance_now_pytorch

A PyTorch Implementation of "Everybody Dance Now" from Berkeley AI lab.
GNU Affero General Public License v3.0
280 stars 72 forks source link

Missing cord in some video #12

Open yijun-li-20 opened 4 years ago

yijun-li-20 commented 4 years ago

Thank you for your work! I have downloaded a Bruno Mars video and resize it to 288*512, but I can't correctly get the coordinates with ./pose_estimator and I found there is one point missing in the results.(Figure 1) I have tried the Afrobeat exercise video and I am pretty sure I get the right setup. (Figure 2) 0201 00000 Source video: https://www.youtube.com/watch?v=PMivT7MJ41M

It would be great if you can check this out!

yijun-li-20 commented 4 years ago

Well I have found the problem of missing joints is not unusual. Can you share how to deal with the -1, -1 after pose estimation?

Lotayou commented 4 years ago

Occlusion is inevitable in real-world videos, and unfortunately this module outputs (-1, -1) for missing cords. Just ignore this and use the rendered figure map as input. It's not a big deal. Alternatively, you can try to use segmentations masks, body parsing maps, or even Densepose to get the pose information.

kai0122 commented 3 years ago

@yijun-li-20 Did you generate a good result when using this target video? I tried on my video for walking pose and also your Bruno Mars dancing pose, but it seems that the result is not good.