Efficient Perceptual Region Detector Based on Object Boundary

Finding nearly semantic ‘visual units’ which visual analysis can operate on is a long-term hard work in computer vision community. Established powerful methodologies such as SIFT, BRISK often extract numerous redundant single keypoints with little information about semantic contents. We propose a novel method called Contour-Aware Regions detector (CAR) to find representative regions in images. Inspired by the recent research conclusion of general object proposal methods that contour is important in object localization, we first alleviate the problem of super pixle overlapping multi-object regions. And then perceptual regions are generated during the merging process using the data structure similar to MSER. Extensive experiments demonstrate: (1) superpixels generated by our method significantly outperform the state-of-out methods, as measured by boundary recall and under-segmentation error. (2) our method can find less and meaningful regions in 0.125 s per image, meanwhile achieve promising repeatabiliy.