A Rotation and Scale Invariant Approach for Dense Wide Baseline Matching

Abstract

This paper proposes a new approach for dense matching of uncalibrated image pair with significant rotation and scale changes. In this approach, a modified region-based matching algorithm is combined with local invariant features like SIFT to conduct dense and reliable matching. First, sparse key point correspondences are established as reference matches; then, in dense matching step, the shape and location of support windows are normalized using SIFT structure information of those reference matches. Thus, scale and rotation changes of input images can be well handled. Experimental results from real data demonstrate that our approach can establish dense and accurate matching in wide-baseline case, which is robust to geometric transformations such as change in scale and rotation, as well as some extent of viewpoint change.