Data augmentation?

[jessicametzger]

I have a general question/feature request about data augmentation. I've been manually adding in additional data augmentation (through the Albumentations package). Applying random photometric distortions (e.g. Gaussian blur, color jitter, etc.) significantly helps with the domain generalization. However, after I started adding random geometric distortions (e.g. random affine distortions), performance on the validation target domain decreased by about 5%, rather than improving, as I've observed in supervised learning. I'm fairly certain this isn't a bug, as I've checked and tested the various pipelines many times. I was wondering, do these types of augmentations mess up the discrimator fooling process? Or, is there anywhere I can find code that adds data augmentation to the fcos training routine?

Thank you!