Image match using distribution of colorful SIFT

Finding reliable correspondence in two or more images remains a difficult and critical step in many computer vision tasks. The performance of descriptors determines the matching results directly. Compared with other descriptors, the Scale Invariant Feature Transform (SIFT) has been used widely for its superiority in invariant attributes, while it will fail in the case of locally visual aliasing. To reduce the perceptual alias of features easily confused, we propose an approach which combines a modified feature descriptor with a novel matching strategy. The feature descriptor is modified by augmenting traditional SIFT vector with dominant hue histogram. A novel matching strategy is developed to validate true matches by establishing geometrical relationships between candidate matching features. The proposed method is tested on many image pairs with viewpoint changes. Based on three instances of geometrical constraint metrics and color information, satisfactory results are attained.