These are rotating face masks, the left calculated for proper lighting conditions, the right filmed from an actual object. Note that the ‘hollow’; (negative) version of the face immediately switches back into a positive view, one cannot ‘hold’; it.
Positive and negative versions of the mask only differ in the position of the assumed light source, and a face is such a strong percept that it overcomes the “history” which is the only cue that the face is negative. On repeated requests, here is my attempt at a more detailed explanation:
First, let’s realise that information of the 3D world is lost when projected on our 2D retinas. So our perceptual system has to reconstruct this, and while this reconstruction attempts get most things right (relying on prior knowledge of the world – possibly using a Bayesian approach) there are retinal images that can be interpreted in more than one way (e.g. any silhouette). The hollow face is a case in point: if we cannot rely on shadows (and in the computer images above we have, of course, different lighting conditions than in your room), there is nothing that can tell us if the face is really hollow or normal.
Second, faces have a special relevance for us; throughout our whole life we try to “read” faces. There are specialised brain areas for faces (fusiform gyrus), and the disease prosopagnosia which occurs from lesions there is a specific loss to recognise faces.
Putting the two facts above together: when both the “hollow” and the “normal” interpretation are equally likely, our sophisticated face processing kicks in and tips the balance toward the “normal” face, since it is trained on such. This obviously does not take history into account, namely the knowledge from the previous rotation angles that the face is hollow. But cognitively we know that it should be hollow – thus arises this strong and beautiful phenomenon.
2009-07-15 The above explanation suggests that it should be stronger for more familiar faces/forms. Indeed, Hill & Johnston (2007) found that it is stronger than veridical stereo information for more familiar forms. Further, adding noise to faces reduces the tendency for depth inversion (see pertinant article on Cognitive Daily).
Roger Garrett pointed out: If you concentrate on the rod that supports the mask in the right movie, you may be able to ‘hold’ the inverted view, at least a bit longer.
This phenomenon is related to “reverspective” and the cute little dragon from Binary Arts, based on a Jerry Andrus design to celebrate Gathering for Gardner 3 (16–18 Jan 1998 in Atlanta, Georgia).
A commentator (thanks!) recently made me aware of this sculpture. Can only be appreciated if you stand in front of it and move sideways, of course – no photo can bring out the phenomenon.