Open julyiii opened 1 year ago
We follow the prior works (e.g. IntagHand) to regress the 3D joints from mano_layer. However, I've also tried using the annotations provided by InterHand dataset and it did not show significant differences in training actually. The possible reason for the difference in evaluation is more likely the GT 3d annotations will have 'valid'
properties to eliminate some joints while this does not exist if we regress from mano_layer.
Thank you for your reply ! Regarding the “valid” annotation you mentioned, I found that different papers have different practices. The first kind of practice is that they don't use the GT 3d "valid" annotations to eliminate some joints in evaluation.(e.g. Intaghand). But The other kind of practice is that they just use corresponding GT 3d "valid" annotations for 3d joints from mano_layer.(e.g. kypt_transformer and interhand). I want to know your practice for fair comparison.
When will the training code be released?