Adapt predictions for temporal consistency

[nito5191]

Hi, thanks for an interesting project and model!

When trying this on video sequences, I sometimes see some noisy/incorrect poses (eg for occluded arms) for a few frames and generally noisy z-coordinates that would be nice to filter temporally to make smoother and less susceptible to outliers. While processing the predicted pose, beta and translation is possible I believe a better solution would be to modify the test-time augmentation and up-weigh temporally consistent poses, which should be possible if I understood the code correctly. For that to work with the latest model I think I would need the raw pytorch model file (non-torchscript), is that possible to get?

[isarandi]

I have some temporal filtering code in another repo here: https://github.com/isarandi/nlf-pipeline
That takes into account the estimated pointwise uncertainties and applies a weighted temporal geometric median-like filter and can fit shared betas over time. In my experience, this fixes most of the outlier/jitter problems.