Closed chingswy closed 1 year ago
Since the dataset is originally used to test the accuracy of multiview human reconstruction algorithms, the smplx models are fit simply by detecting keypoints from 6 views. Thus, in some multiperson cases the fitting may not be accurate. We don't use ICP or markerless mocap. But if you want to get more accurate results, surely you can include ICP and many other algorithms.
Thanks!
I visualized
MultiHuman/three/[500,502]
sequence, it seems that some body parts of smplx are misaligned with the textured mesh. Did you use ICP or markerless mocap to get the SMPLX results?