I argue that in spite of the progress in several machine vision applications, the general machine vision problem is not going to be solved any time soon. There are three reasons for that: (1) The complexity of human vision: Bottom-up and topdown processes are tightly interwoven and we have no good models for dealing with that; (2) The fact that perceptual similarity is not the same as mathematical similarity; (3) The illusion of progress by relying on “proofs by example” that are not always valid. I discuss several examples of applications that were successful because they did not face any of the three obstacles.

Paper in PDF format (173KB)

* To appear in the Journal of Signal, Image and Video Processing, January 2014 (Invited Paper). The journal is published by Springer and the paper has been already posted in their web site.
Link to Springer site