IdentifyNet for Non-Maximum Suppression

Two-stage object detectors have achieved great success in recent years. However, recent work mostly focuses on optimizing loss functions or learning multi-level feature representation, while introducing additional homogeneous task to improve detection has been under-explored. In this paper, a novel framework named as IdentifyNet is proposed, which incorporates an additional identification task to enhance the feature learning of region proposals. Specifically, besides classification and bounding box regression, the proposed IdentifyNet further learns to predict whether two different region proposals belong to the same object, thus forcing the network to learn more informative and representative features for different proposals, especially for those from the same object class. Moreover, current detectors apply greedy non-maximum suppression to remove duplicated boxes whenever their Intersaction-over-Union (IoU) exceeds a preset threshold, which would fail when two boxes largely overlap with each other while belonging to two different objects of the same class. To overcome this, we further propose a novel decode non-maximum suppression algorithm by taking advantage of the predicted identity information of different proposals from the identification task. Extensive experiments on PASCAL VOC 2007, VOC 2012 well demonstrate the proposed method can greatly improve detection performance.