search

hongsukchoi / TCMR_RELEASE

Official Pytorch implementation of "Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video", CVPR 2021
MIT License
274 stars 39 forks source link

strange 3D #1

Open potie1013 opened 3 years ago

potie1013 commented 3 years ago

image I tried your model, but found some frames' result very strange, just like this image, need your help

hongsukchoi commented 3 years ago

Is this frame near the end of the input video? (1~3 second before the end) I found some weird phenomenon like this in our demo code recently and trying to figure out the reason.

In other cases, this could be due to motion blur in near frames.

giuli13 commented 2 years ago

I have the same problem when I pass the joints2d groundtruth to have a better mesh fitting. I my case the wrong mesh fitting is not near the end of the input video.

hongsukchoi commented 2 years ago

Hmm... I remember VIBE's demo implementation had some post-processing using GT 2d joints, but I think I removed those codes in this repo. Did you try without GT 2d joints and got the same strange result? You could also try another pretrained models https://drive.google.com/drive/u/1/folders/1a_nv8MfvToutatBvKaIXP7J_7U4BjSu2

hongsukchoi commented 2 years ago

@giuli13

The pretrained model is trained on 25~30fps videos. So I set the demo code to read video by 30 frames per second. If your video originally has low fps, the number of frames will increase and the image qualities will decrease than what you know. This might have affected the accuracy. You can try to remove the frame rate line as here and run the demo code