UFLDL for Shape Constancy

I was wondering if anyone knows of any past/present/future research that applies unsupervised feature learning/sparse coding to teach an agent to detect shape constancy? By shape constancy I mean recognizing an object or at least the shape of an object as the subject's viewing angle changes, for instance, 2D and 3D panning.

I arrived at this question having just watched all of a brief talk by Dr. Andrew Ng (http://www.youtube.com/watch?feature=player_embedded&v=ZmNOAtZIgIk). In the middle-chunk of his presentation he gives a simplified explanation of a sparsely-coded neural network approach to learning that has been applied to images, audio, and video where the agent subjectively learns which features to detect and encode. I wonder if the same kind of algorithm would be able to recognize a shape as it is viewed at an increasingly sharp angle (in which case the shape becomes a shrinking sliver)? Thanks!