Channel–Spatial fusion aware net for accurate and fast object Detection

Abstract

A major challenge of object detection is that accurate detector is limited by speed due to enormous network, while the lightweight detector can reach real-time but its weak representation ability leads to the expense of accuracy. To overcome the issue, we propose a channel–spatial fusion awareness module (CSFA) to improve the accuracy by enhancing the feature representation of network at the negligible cost of complexity. Given a feature map, our method exploits two parts sequentially, channel awareness and spatial awareness, to reconstruct feature map without deepening the network. Because of the property of CSFA for easy integrating into any layer of CNN architectures, we assemble this module into ResNet-18 and DLA-34 in CenterNet to form a CSFA detector. Results consistently show that CSFA-Net runs in a fairly fast speed, and achieves state-of-the-art, i.e., mAP of 81.12% on VOC2007 and AP of 43.2% on COCO.