Learning Contrastive Representation for Semantic Correspondence

Dense correspondence across semantically related images has been extensively studied, but still faces two challenges: 1) large variations in appearance, scale and pose exist even for objects from the same category, 2) labeling pixel-level dense correspondences is labor intensive infeasible to scale. Most existing methods focus on designing various matching modules using fully-supervised ImageNet pretrained networks. On other hand, while a variety of self-supervised approaches are proposed explicitly measure image-level similarities, pixel level remains under-explored. In this work, we propose multi-level contrastive learning approach semantic matching, which does not rely any model. We show that key component encourage convolutional features find between similar objects, performance can be further enhanced by regularizing cross-instance cycle-consistency at intermediate feature levels. Experimental results PF-PASCAL, PF-WILLOW, SPair-71k benchmark datasets demonstrate our method performs favorably against state-of-the-art approaches.