Task differentiation: Constructing robust branches for precise object detection

Abstract

Most prevailing object detection methods share networks and features between localization and classification components, which easily leads to sub-optimal learning for the two separate tasks. In this paper, we propose a conception of task differentiation and design specialized sub-networks for both localization and classification components based on SSD framework. A novel probability based localization method is introduced into the one-stage framework and combined with bounding box regression for precise object localization. Furthermore, a new feature fusion strategy, together with a global attention mechanism, is proposed to learn more robust features. Experimental results on PASCAL VOC and MS COCO data sets indicate that our method has impressive performance compared with other state-of-the-art object detection approaches.