Abstract
We propose that it is worth looking at intelligent brute force methods for machine vision with an eye towards figuring out what can be done with tera-to peta- ops/byte resources. In this paper, we work with an object recognition system that uses an intelligent, brute-force approach, and achieves results for the problem of general 3D object recognition in clutter that are as good or better than anything demonstrated to date. It does not scale to human-level performance, or use peta-op resources (yet). However, the performance of this system, and of a few other recently reported resource intensive systems, is good enough to allow us to begin investigating other aspects of human visual intelligence. In particular, we look at the problem of acquiring visual recognition models from unlabeled imagery. Some of our experiments, making extensive use of an already resource intensive recognition system, do utilize approximately 1015 operations (a peta-op).