Improve Semantic Correspondence by Filtering the Correlation Scores in both Image Space and Hough Space

Abstract

Semantic correspondence is an important and challenging task in computer vision due to background clutter, especially for analyzing images from the same category but with large intra-class variations. The methods using regularized Hough matching only consider global offset consensus to enhance geometric consistency, where the filtering in Hough space is usually used to aggregate the scores of the similar offsets. However, the matched objects between two images are usually located in a local region of the images. Therefore, in this paper, we propose an improved method for establishing semantic correspondence by introducing regional offset consensus, which is implemented by filtering the correlation scores in both Hough space and image space, where the filtering in image space is used to enforce the regional offset consensus. Experiments show that the proposed method can produce superior results than the state-of-the-art methods, when applied on four benchmark datasets including a large-scale SPair-71k dataset.