Object locating using a single model view

We address the problem of locating a 3D object in a cluttered image using a single clean image of the object as the model. We represent an object as a graph, with the high-curvature points in its image as the graph nodes and the low-curvature contours joining them as the graph arcs. We propose a mechanism which is based on a geometric constraint we call the flatness constraint, that would allow many node correspondences to be extrapolated from four of them. We point out exactly what class of objects would admit a full determination of node correspondences under the mechanism. We also describe how we get the first four node correspondences, and how we use the node correspondences to construct the arc correspondences. The integral of the image gradient magnitude along the arcs in the scene provides a convenient measure on how good the scene object is matched to the model object