A Lexical Analogy to Feature Matching and Pose Estimation

We relate the problem of finding a correspondence between sensed and model features to that of finding a match between a random set of letters and words in a dictionary. The process is equivalent to hashing and the lexical perspective illuminates items such as design tradeoffs, computational complexity, and hashing function definition. A method for two-dimensional pose estimation based on this concept has been implemented. The method is local feature based and is robust to image warping, occlusion, illumination anomalies, and sensed feature generation errors. The method will work with certain modifications for threedimensional data. The domain is restricted to translation and rotation invariant applications, since many pose estimation problems do not require scale and skew invariance. This non-affine constraint can reduce computational and storage complexity vis-à-vis a fully affine transformation invariant technique. Keywords—pose estimation, hashing, hash tables, lexical, computer vision, featurematching, feature correspondence, pose clustering