Open LightOfMonet opened 1 year ago
Please refer to #94.
ok! thanks for the great work
When I was recreate your work, my PAMPJPE was larger than 60, I think the data on my side was not aligned. Can I ask you a few questions?
How did you deal with Human3.6m and 3DPW? I see other works that provides annotations for native coordinate systems, such as SPIN and PyMAF. Can I just use their label here, and then convert smpl directly to f=√w2 + h2? By the following:
Or I have to compute global_t, like global_t in the Pseudo-GT you provided?
The f of 3DPW training set is calibrated f and its global_t is calculated. Can I still supervise 2d reprojection loss and smpl 3d loss with f=√w2 + h2?
Is it necessary to supervise global_t? Is global_t supervised after cam_crop2full ()?
Where is the testing code? How can I trust the sota values of your model?