You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thank you for this project. It is very helpful. It seems to have a discrepancy in the normalization of the image. I noticed two potential issues:
The detector retinaNnet applies some normalization which also affects the face crop data.face. Then, the normalization order should stay the same in the unifier to denormalize the face crop into [0,1]. Right now the face crop can have a negative value.
The image is loaded with torchvision resulting in RGB images. From the RetinaNet implementation preprocessing the image is loaded in BGR format the mean they applied is (104, 117, 123). In facetorch, the transform is applied before rgbtobgr() thus I think that the correct order would be (123,117,104).
With the modification of point 1, all face crops are in [0,1]. For point 2 it is harder to check
The text was updated successfully, but these errors were encountered:
Hello,
Thank you for this project. It is very helpful. It seems to have a discrepancy in the normalization of the image. I noticed two potential issues:
With the modification of point 1, all face crops are in [0,1]. For point 2 it is harder to check
The text was updated successfully, but these errors were encountered: