Resulting shape and pose is not good. about muvs

[ZhiweiREN97]

Hi @abdullahchand, I also have the same problem. Actually, if I used the camera parameters the author provided on my own video, the results is OK. But if I inputted my camera parameters, the results are poor for most of the frames. I guess there might be some issues about the format of the camera parameters. I noticed that in the camera.py the T is divided by 1000, I multiply it on my own parameters but it doesn't work.

Would you mind telling me if you solved this problem?

[abdullahchand]

Hi @ZhiweiREN97 . I actually I was able to solve it for most of the videos. I was able to get the intrinsic parameters correct and used the extrinsic parameters from the default camera.mat. They worked well but not for all Videos. In SMPLfy the authors predict the extrinsic parameters through fitting the model and only need the extrinsic parameters in there. The problem comes buecause in this repository, that feature has been replaced by absolute parameters. We have to give extrinsic and intrinsic parameters.

[ZhiweiREN97]

Thanks @abdullahchand! I absolutely agree with you. I also found out that if I use my own intrinsic parameters and keep the default exterinsic parameters unchanged, the results are much better than I use both intrinsic and extrinsic parameters of my own. However, I need to projected the mesh back to the images which is not good enough if the extrinsic parameters are not changed. I guess the problem is the bad initial parameters. I would update it if I found a perfect solution to that.

[ZhiweiREN97]

Hi @abdullahchand! I found a temporary solution for this. You can use your own camera parameters and you need to multiply the "t" by 1000 (at least for me that works). Then, you may choose two views with a good deepcut result. The loss get to 10e2 for me which is much better than before.