Gaussian Aware Anchor-Free Rotated Detector for Aerial Object Detection

Object detection in specific scenarios has received increasing attention for many applications. However, the feature inconsistencies of localization and classification branches in remote sensing models may degrade the detection performance. Furthermore, the existing IoU-based label assignment strategy cannot accurately capture objects' shape and oriented information. We propose an anchor-free detector called the Gaussian aware rotated detector (GARDet) to address the above issues. It contains two improvements: the feature alignment module (FAM) and the Gaussian dynamic label assignment (GDLA) strategy. FAM consists of oriented feature alignment (OFA) convolutions sensitive to orientation-invariant features inside objects and spatial feature alignment convolutions sensitive to spatial coordinate information. GDLA uses a Gaussian matching confidence (GMC) based on the Gaussian distance to measure the quality of the predicted bounding boxes and dynamically assigns positive and negative samples for training. Extensive experiments on remote sensing object detection datasets (DOTAv1.0 and HRSC2016) demonstrate that the proposed model can achieve competitive performance.