Feature-Aware Network Based on Upper Triangular Interaction for Single Shot Detector

Dong, Xiang Li, Feng Bai, Huihui Zhao, YaoConsidering the impact of the balance between real-time performance and detection accuracy, single-stage detectors have received more attention than two-stage detectors. However, due to a large number of detection objects with different sizes, the existing detectors still have limitations in terms of scale variations. To improve the situation, we propose a feature-aware network based on upper triangular interaction (UTINet), which introduces an upper triangular interaction module (UTIM) to focus on features of different scales and progressively interact with each other. Moreover, we build a feature-aware enhancement module (FAEM) to learn more comprehensive feature representations by amplifying the feature differences before and after fusion. Experiments are performed on MS COCO, and our UTINet with a $$300 \times 300$$ input achieves superior performance of 33.3% mAP compared with existing methods.