Learning Rich Detailed Features and Object Relations for Object Detection

This paper proposes a feature fusion method with multi-scale detail features and an object relation inference method to better assist the detectors in completing detection tasks. First, we fully use and extract the multi-scale detail features contained in the shallow feature map of the backbone, and integrate them into the feature map of the corresponding scale responsible for detection, to realize the effective transfer and supplement of detail features. Secondly, we fully use the location relationship between object categories for inference and judgment so that the model can more fully use the potential relationship between objects to assist detection instead of using the features of the object itself. Finally, we conducted extensive experiments using different detectors and backbones as the improvement baseline to verify the effectiveness of our proposed method. The results show that our improved model significantly improved the prediction accuracy compared with the baseline model.