Problems with existing image classifiers.

https://www.youtube.com/watch?v=sZViVi1rZPA

What this is probably showing is that vision systems naively trained are learning absolute pixel values or color histograms, whereas biological vision systems are mainly concerned with edge detection and optical flow.

Probably running the input image through a Canny edge detector, or using that as additional input during training, would help to give a more robust result.