FCOS-Lite: An Efficient Anchor-free Network for Real-time Object Detection

Fully Convolutional One-Stage detector(FCOS) is a simple and strong anchor-free detector. However, the outstanding performance comes at the cost of long inference time. In this paper, a more efficient anchor-free detector FCOS-Lite is proposed to speed up the inference time of FCOS. FCOS-Lite improves the three components of FCOS, namely the backbone, neck and head. We first apply an efficient attention module in the backbone to extract more important semantic information. Then, we propose an adaptive feature-fusion module in the neck to detect small objects accurately. Finally, we use some strategies in the head of FCOS-Lite to reduce the computation. With these methods on the FCOS-Lite, we achieve a speed-accuracy balance on the MS COCO dataset. FCOS-Lite's inference time is almost halved based on the FCOS with a slight accuracy drop. (33.2% mAP at 24 ms for FCOS-Lite compared to 37.4% mAP at 44ms for FCOS on COCO). FCOS-Lite suits for the real-time object detection, improving both accuracy and efficiency of the real-time detector YOLOv3 and FCOS-mobileNet(31.0% mAP at 29ms for YOLOv3 and 30.9% mAP at 27ms for FCOS-mobileNet on COCO).