A Double Stream Module in Backbone for Object Detection Network

The representation ability of the network is crucial for object detection. In this paper, we present a novel building block named double stream block to construct a backbone called double stream module (DSM) to improve the representation ability of object detection networks. The double stream block is designed by fusing the shuffle branch and the non shuffle branch with a weight layer. To make our block adaptive for the depth of the convolutional layer, a shortcut connection is adopted after the double stream fusion. After that, DSM is built as a backbone and integrated into Faster-RCNN, Cascade-RCNN, DetectoRS detection system to verify its applicability. On the VisDrone dataset of VAV, the evaluation results for object detection show that the double stream design can significantly increase the representation ability of the network, and our approach gains 2.0, 0.5and 1.1 mAP improvements in the detection systems mentioned above respectively.