Recognizing Flexible Objects

Pedro F. Felzenszwalb
Daniel P. Huttenlocher

Abstract

In this paper we present an approach to the problem of recognizing objects using generic, flexible, parts-based models. Our approach uses MAP estimation to find the globally optimal location of the model parts in an image. We identify an interesting class of models where the dependencies among the parts form a tree. The key result is that for such models the MAP estimate can be found efficiently using dynamic programming and a novel generalization of distance transforms. This distance transform method is useful in its own right, for a fairly broad range of minimization problems. We illustrate our approach using a generic model of a person, and show results of finding people in complex scenes under a broad range of lighting conditions and in varied poses. The method is quite practical, our implementation runs in just a few seconds on a desktop workstation.

Click here for a paper describing the work in detail (postscript, gzip).


Person model and recognition result

recog2.jpg (31894 bytes)

Click here for more example results.


Back to Vision Homepage