An Initial Study of an Adaptive Hierarchical Vision System

Abstract

We describe an empirical study of an adaptive, hierarchical vision system. Using a simple vision task requiring both low-level and high-level processing, we examined how three schemes of feedback for on-line learning affected the true positive rate, the number of instances used for learning, and the need for user feedback. The rst scheme used for learning those instances for which the user provided feedback. The second used all instances, assuming that no feedback meant correct classication. In the end, a hybrid scheme with each method at dierent levels yielded the best results, showing that more examples for learning signicantly improved the true positive rate of the classiers at the lower level, but not at the higher level. Furthermore, this hybrid method did not increase the need for user feedback. 1. Introduction Hierarchical vision systems (Mohan & Nevatia, 1989) work by repeatedly grouping visual constructs into higher-level constructs and selecting the m...