Randomized Global Transformation Approach for Dense Correspondence

Motivation Recently, many researchers have begun to attempt to solve dense correspondence problem for more challenging images which have high variability in terms of photometric and/or geometric conditions [4]. For these challenging scenarios, there exist two principal bottlenecks which make conventional methods provide limited performances; (1) photometric variations derived from different camera specifications, illumination or exposure conditions, and (2) geometric variations derived from viewpoint changes, object pose changes, and non-rigid deformation for objects [1]. For the first bottleneck, many robust feature descriptors have been proposed to alleviate photometric variations. As a pioneering work, the SIFT flow [2] has shown satisfactory results on different image pairs having semantically similar property by employing the SIFT descriptor. However, geometric variations for the second bottleneck still remain unsolved due to its large search space, including translation (or flow), rotation, and scale, which induces extremely large computational complexity. In this paper, we propose a randomized global transformation approach to estimate reliable correspondence between challenging image pairs having photometric and geometric deformation. Our approach starts from an intuition that geometric variations between two images can be formulated as piecewise transformation model as shown in Fig. 1. Finding the optimal global transformation for each pixel enables us to infer correspondence fields without computing geometric fields.