Closed Fouksi closed 1 month ago
Thanks! We used both EFT and SPIN fits for experiments in our paper.
The released code uses EFT fits as training with them performs slightly better. We did not release the EFT fits for MPII since training only with H36M, MPI-INF-3DPH and COCO had the same performance. See here for the datasets and their ratios used during training -- ratios are the ones proposed in PARE.
For experiments with SPIN fits we used the datasets (which include MPII) and ratios as seen here to fairly compare with ProHMR that also uses SPIN fits during training. You can get the SPIN fits for all datasets, including MPII by following the instructions here.
And where can I find the validation set for COCO-EFT? The npz file you supplied doesn't have the pose & shape parameters. Thanks.
We only used the 2D keypoints from the validation set of COCO for monitoring purposes during training, so I don't have an npz file with the COCO-val EFT fits.
You can check out the official EFT repo if you need the pseudo-GT for some reason.
Great paper!
I can't find your MPII annotations - in which method did you use to annotate it?
Thank you.