Spatial Attention for Multi-Scale Feature Refinement for Object Detection

Scale variation is one of the primary challenges in the object detection, existing in both inter-class and intra-class instances, especially on the drone platform. The latest methods focus on feature pyramid for detecting objects at different scales. In this work, we propose two techniques to refine multi-scale features for detecting various-scale instances in FPN-based Network. A Receptive Field Expansion Block (RFEB) is designed to increase the receptive field size for high-level semantic features, then the generated features are passed through a Spatial-Refinement Module (SRM) to repair the spatial details of multi-scale objects in images before summation by the lateral connection. To evaluate its effectiveness, we conduct experiments on VisDrone-2019 benchmark dataset and achieve impressive improvement. Meanwhile, results on PASCAL VOC and MS COCO datasets show that our model is able to reach the competitive performance.