Closed Darius-H closed 10 months ago
Hi,
Thanks for bringing this up. I used to store the masked images in the masks_sam_images
folder to verify if the segmentation is correct. As we are not using these images for training it shouldn't influence the results.
Regarding the keypoints, yes, we can actually switch to that easily by adding --hand
here.
Best, Tianjian
https://github.com/tijiang13/InstantAvatar/blob/4348c1e5a38627a8096d288072896e281577d4f5/scripts/custom/run-sam.py#L26C1-L37C68 dirname in makedirs is "masks_sam_images" but the last line in replace is "masked_sam_images".
And i found this segmentation model struggles in segmenting human hand from the background. Maybe adding hand keypoints will be better.