Downscaling vs cropping ROI masks for CNN

[ethanbb]

I have recently been sifting through a ton of very elongated ROIs, due to ambient light from our projector that we sometimes weren't able to entirely block. Just a thought - it seems like the CNN should be able to automatically reject them. I believe that right now it's not because the mask image is cropped to a size related to gSig before passing it in.

Would scaling the nonzero portion of the image down rather than cropping it help? Or more likely the region within a certain %ile or energy threshold. I realize this could be tricky and I'm wondering if it's been considered/tried in the past. Ideally this would also reduce the awkward dependence on gSig that sometimes necessitates using the CNN multiple times at different sizes. (Though maybe the size itself should be another criterion then).

Alternatively, for elongated ROIs specifically, it looks like there was an eccentricity metric in the past, but it doesn't seem to be well-documented now - has anyone used that recently/know if it still works?