When running Sapiens on the released test sequence to obtain 2D poses, the predictions appear noisy and mostly inaccurate (mislocalized joints) compared to released skeleton. I’d like to confirm expected performance on this data and ask the following:
(1) fine-tuning: Have you fine-tuned Sapiens for this specific released test sequence or a similar domain?
(2) Alternative 2D estimator: Is it advisable to plug in a different D pose estimator that outputs fewer keypoints than Sapiens produce?
When running Sapiens on the released test sequence to obtain 2D poses, the predictions appear noisy and mostly inaccurate (mislocalized joints) compared to released skeleton. I’d like to confirm expected performance on this data and ask the following:
(1) fine-tuning: Have you fine-tuned Sapiens for this specific released test sequence or a similar domain?
(2) Alternative 2D estimator: Is it advisable to plug in a different D pose estimator that outputs fewer keypoints than Sapiens produce?