Back-Projective Priming: Toward Efficient 3d Model-based Object Recognition via Preemptive Top-down Constraints

Abstract

This thesis introduces back-projective priming, a computer vision technique that synergistically fuses object recognition and pose estimation by augmenting 3D models with geometric constraints. It also enables the use of image features too indistinct for use by other model fitting algorithms such as geometric hashing. To efficiently accommodate features that do not provide a scale attribute, we've developed a "match pair" finding heuristic called second-order similarity that reduces model fitting time complexity from a worst case of O(N^2) to O(N*Log(N)). An object recognition problem that is simple, practical, and well explored by other researchers is the problem of locating electrical outlets from the vantage point of a mobile robot. To demonstrate the relative merits of back-projective priming, we use it to build a system capable of locating generic electrical outlets in unmapped environments. Compared to our baseline algorithm, back-projective priming is shown to provide superior sensitivity when dealing with the challenges of low contrast, perspective distortion, partial occlusion, and decoys.