Visual Similarity, Judgmental Certainty and Stereo Correspondence

Abstract

Normal human vision is nearly infallible in modeling the visually sensed physical environment in which it evolved. In contrast, most currently available computer vision systems fall far short of human performance in this task, and further, they are generally not capable of being able to assert the correctness of their judgments. In computerized stereo matching systems, correctness of the similarity/identity-matching is almost never guaranteed. In this paper, we explore the question of the extent to which judgments of similarity/identity can be made essentially error-free in support of obtaining a relatively dense depth model of a natural outdoor scene. We argue for the necessity of simultaneously producing a crude scene-specific semantic "overlay". For our experiments, we designed a wavelet-based stereo matching algorithm and use "classification-trees" to create a primitive semantic overlay of the scene. A series of mutually independent filters has been designed and implemented based on the study of different error sources. Photometric appearance, camera imaging geometry and scene constraints are utilized in these filters. When tested on different sets of stereo images, our system has demonstrated above 97 correctness on asserted matches. Finally, we provide a principled basis for relatively dense depth recovery.