Alignment by Composition

We propose an unsupervised method to establish dense semantic correspondences between images depicting different instances of the same object category. We posit that alignment is compositional in nature and requires the detection of a similar visual concept between images. We realize this in a top-down fashion using objectness, saliency, and visual similarity cues to co-localize the regions of holistic foreground objects. Jointly maximizing visual similarity and bounding the geometric distortion induced by their configuration, the target foreground object is then composed by the subregions of the source foreground object. The resultant composition is used to form a dense motion field enabling the alignment. Experimental results on several benchmark datasets support the efficacy of the proposed method.