Articulated Matching with Point

Articulated matching arises naturally in object recognition and tracking. When a part-based description of an object is adopted, the problem of matching the model features to the data can be decomposed into matching the parts while paying attention to overall model coherence. Here we focus on the rst part of the problem, namely, the articulated matching of a part based description of an object to unlabelled data point features. Each model part is assumed to undergo an independent aane transformation. The novel aspect of our approach is the integrated search for both the point-to-point feature correspondences and the diierent aane transformations of the parts. The result is an alternating algorithm which uses a least-squares solution for the aane parameters and the softassign technique for the correspondences. A signiicant property of the algorithm is its ability to reject missing and extra features as outliers. The algorithm is embedded in a deterministic annealing scheme. Results are shown on synthetic and real-world imagery.