Contextual disambiguation for multi-class object detection

We consider the problem of detecting and localizing instances from multiple object classes. Suppose an overcomplete index - an initial list with extra detections but none missed - is provided. We and others have previously shown how this can be done efficiently with coarse-to-fine search. How would one prune such a list to a final interpretation? We propose a method based on contextual disambiguation: first, a Viterbi algorithm is utilized to extract N candidate interpretations by using the global context to provide constraints among object classes or poses. Then, the extracted candidates are compared in a pairwise fashion to resolve remaining ambiguities, and the final interpretation is constructed. The whole procedure is illustrated by experiments in reading license plates.