It seems like EXIF data isn't used enough in machine learning; for instance if you want your model to know what color something "actually" is, it would help to feed it the white balance and exposure values, not just the pixels.

(Doesn't help they're all trained on low dynamic range sRGB images and not real camera data.)