Model-Based Pose Proposal for 2-D Object Recognition

We consider the problem of nding a known two-dimensional object in an image, or verifying that it does not appear in the image. We adopt the strategy of doing a fast scan for potential places in the image where the object could be; we call this scan pose proposal. Each pose hypothesis is a set of edges that correspond to a subset of the transformed object boundary. Our algorithm works by nding U-shaped segments of object boundaries, doing a quick match process between U-shaped segments in the image and the model, and combining the matches into overall pose hypotheses. Analysis and experiments show that the algorithm runs eeciently, and does a good job of discarding all but a few spots in the image as possible pose hypotheses.