It would be fairly easy to construct something like this it's never seen before to test it. Go try it. Cut some shapes out of paper or 3d print something weird.

It would be fairly easy to construct something like this it's never seen before to test it. Go try it. Cut some shapes out of paper or 3d print something weird. I'd suspect that a good model will infer what it can from universal depth cues (E.g. Bokeh and perspective) when present but not perform as well if they're as not there or as well as it does on familiar objects.