On image correspondence using topology preserving mappings

A computational approach for establishing correspondence between two image views is presented. We show that a self-organizing feature map trained with tokens (features) from the first frame and subsequently with tokens from the second frame (without re-initialization) is capable of indicating the underlying transformation which results in the second frame. The reliance of the self-organizing feature map on the underlying probability density of the features makes the proposed approach insensitive to missing tokens. Simulation results under three different observer movements (panning, zooming, and rotation) are presented to illustrate the proposed method