Can't reproduce results on Human3.6 (Table 1)

Hi, thanks for the good job.

I tried to run:

python train/train_prohmr.py --root_dir=prohmr_reproduce/

to re-trained the model myself. By evaluating the last checkpoint (iteration=500k), I got the performance:

Screen Shot 2022-10-12 at 6 21 39 PM

However, the PA-MPJPE reported in the Table 1 of the paper is 41.2, which is lower than 42.8 as I got. Moreover, the evaluation from the given data/checkpoint.pt is

Screen Shot 2022-10-12 at 6 15 37 PM

which is actually better than the reported number and thus shows a larger gap with my re-trained weights.

Have I missed any details here to reproduce the performance reported in the paper?

nkolot / ProHMR

Can't reproduce results on Human3.6 (Table 1) #24