With a larger picture like 1024*1024, the train results poor loss like 10E6 about muvs

[YinghaoHuang91]

Do you mean that you tried the code on your own dataset? If so you need to update the settings like camera parameters and image size accordingly.

[MisssRain]

Thanks a lot! I figured out that it was our camera's extrinsic parameter, transpose matrix, that has already been divided by 1000.0. The loss function falls down to 10E4 for pose(\thetas) and shape(\betas) parameters in smpl then.
The training result projected to 2d image was not good enough under condition that we have 8 camera views...
Another question: I didn't find where to set the picture size in your code, does it matter for the training? Or if I should change the Data/genericPrior.mat?

[YinghaoHuang91]

You need to dig deep about how camera parameters are represented. I think this link might help you a lot: https://docs.opencv.org/2.4/doc/tutorials/calib3d/camera_calibration/camera_calibration.html. No prior needs updated, I think it's the camera parameters that needs adjusted.

[MisssRain]

Thanks very much! Really a great help!