Low-dimensional superpixel descriptor for visual correspondence estimation in video

Estimating local visual correspondence between video frames is an important and essential challenge in many visual applications. Keypoint based sparse matching is a common way to address the problem of local visual correspondence estimation. This paper proposes a local visual correspondence estimation method based on extracting discriminative features from superpixels. In the proposed approach, superpixels are generated by over segmentation at first. Then the superpixels are described by orientated center-boundary distance (OCB-D) and gray-level co-occurrence matrix (GLCM) which extract shape feature and texture feature, respectively. Experimental results on the widely used Middlebury dataset prove that the proposed superpixel descriptor achieves much higher accuracy than the compact ORB descriptor when same dimensions of features are used. In addition, benefited from its low-dimension character, the proposed descriptor is memory-efficient and hardware friendly.