Cross-scale Dynamic Relation Network for Object Detection

The majority of object detectors only consider the features in region proposals, without taking the global context or the relationships between objects into detection. Conceivably, it would inevitably limit the improvement of performance. To tackle the problem, we introduce a Cross-Scale Dynamic Relation Network (CSDRN) that can explore the relationships between specific objects in an image, and its core components include a Cross-Scale Semantic-Aware Module (CSSAM), Dynamic Relation Graph Reasoning (DRGR), and Semantic Attention Fusion Module(SAFM). Through the CSSAM, the crucial information in feature maps of different scales achieve semantic interaction to obtain a cross-scale semantic feature. We activate the category knowledge in the image and combine the cross-scale semantic feature to create a dynamic relationship graph. Therefore, we can get more precise relation between objects. Guided by the relation, a semantic attention is generated to enrich the visual features. Experimental results on the COCO dataset show that the proposed CSDRN can effectively improve the detection performance, reaching 54.8% box AP, which is 3.9% box AP over the baseline. Moreover, 47.6% mask AP is achieved in instance segmentation, exceeding the baseline 3.6% mask AP.