Directional geometric histogram feature extraction and applications

Abstract

Image feature has been a hot research topic within the field of computer vision, with a wide scope of direct impacts on detection, recognition, image retrieval and pose estimation, etc. In this paper, we propose a novel image feature: Directional Geometric Histogram (DGH) which adopts directional geometric approximation from the geometric Bandelet transform to enhance the description distinctiveness and selectivity among monocular images, particularly by renovating the histogram of geometric regularity to characterize local image context with human objects. Other than the image geometry defined over edges, our approach can well preserve inner and outer patterns of contours with strict geometry. We have compared the proposed method with classic global features and conducted comprehensive experiments in human detection, pose estimation as well as scene recognition tasks on various datasets. Final evaluation results show that the dimensionality of the DGH feature can be reduced to less than half of the original size, which is also sparse while keeping competitive discriminatory effectiveness and distinctiveness in such visual tasks. Besides its relaxed computational requirement and off-the-shelf theoretical backup, the method is in the meanwhile quite promising for potential fields in video surveillance, pattern identification, etc.