Thanks a lot for sharing your code, it really helped!
A quick question regarding the visual objectives (predicting objects/attributes for Faster RCNN regions).
If I'm understanding the loss calculation correctly, the feats_mask only applies to feature regression, not object/attribute prediction, so even if a patch is randomly zeroed-out or being replaced by another patch, it still needs to predict the original labels for objects/attributes.
Thanks a lot for sharing your code, it really helped!
A quick question regarding the visual objectives (predicting objects/attributes for Faster RCNN regions). If I'm understanding the loss calculation correctly, the
feats_mask
only applies to feature regression, not object/attribute prediction, so even if a patch is randomly zeroed-out or being replaced by another patch, it still needs to predict the original labels for objects/attributes.Is that correct? Am I missed something here?
Many thanks!