One-Stage Object Detector Using Feature Fusion and Dual Attention

Multi-scale object detection and small object detection are important tasks in computer vision and robotics. Traditional one-stage object detection algorithms are fast, but cannot adapt to changes in object scale perfectly and have low accuracy for small object detection. In this paper, we proposed a one-stage object detector using feature fusion and attentional mechanisms. This detector has better adaptability and higher accuracy with marginal extra cost. Our algorithm is based on the CenterNet framework, with modifications and enhancements. First, we used a feature pyramid network with bidirectional cross-scale connections to enable the full integration of lower layers and topmost features, which improved the network's adaptability to scale changes. Second, we proposed a dual attention module that integrates spatial attention and channel attention and added it to every layer of the pyramid, so that the network can focus more on useful information, which enhances the network's capability to detect small objects. Using a lightweight backbone, our algorithm achieves 40.2 mAP on COCO at 33 fps on a Titan Xp GPU, up to 3 mAP compared to CenterNet with the same backbone. Besides, for the small targets in COCO, our algorithm achieves 20.63 mAP, Which is 30% higher than CenterNet. We also tested our algorithm on a body part detection dataset of Pascal, Voc,which further demonstrates the effectiveness of our algorithm in terms of application.