RGRN: Relation-aware graph reasoning network for object detection

In the field of object detection, due to the complexity of realistic scenarios, the objects are mostly obscured and semantic-confusable. The existing CNNs-based object detectors focus only on the information within the region proposal and ignore the auxiliary role of objects-objects relationships, leading to difficulties distinguishing difficult samples in complex spaces. Accordingly, in this paper, we propose a novel relation-aware graph reasoning network (RGRN) to adaptively discover and integrate key semantic and spatial relationships in images. Specifically, in order to realize information interaction and relational reasoning between nodes, we design two parallel modules: the semantic relational reasoning module (SRRM) and the spatial relational reasoning module (SPRM). SRRM mines the semantic relationships between objects by discriminating the semantic similarity between graph nodes, and SPRM finds the spatial relationships between objects by the relative positions between nodes. Our method considers the relative spatial location and semantic correlation between objects, which can easily embed in existing networks in real-time to improve performance. Solid experiments verify the effectiveness of our method, which achieves around 16 % improvement on MS COCO and 10 % on PASCAL VOC in terms of mAP and outperforms the state-of-the-art relation-based methods, which indicates the superiority and effectiveness of RGRN.