Deep contour and symmetry scored object proposal
Object proposal has been successfully applied in recent supervised and weakly supervised visual object detection tasks to improve the computational efficiency. The classical grouping-based object proposal approach can produce region proposals with high localization accuracy, but incorporates significant redundancy for the lack of object confidence to evaluate the proposals. In this paper, we propose leveraging the essential properties of images, i.e., contour and symmetry, to score the redundant region proposals. Specifically, the contour and symmetry are extracted by a Simultaneous Contour and Symmetry Detection Network (SCSDN) and used to score the bounding box with a Bayesian framework, which guarantees that the scoring procedure is adaptive to general objects. A subset of high-scored proposals reserves the recall rate, while can also significantly decrease the redundancy. Experimental results show that the proposed approach improves the baseline by increasing the recall rate from 0.87 to 0.89 on the PASCAL VOC 2007 dataset. It also outperforms the state-of-the-art on AUC and uses much fewer object proposals to achieve comparable recall rate.