OBBStacking: An Ensemble Method for Remote Sensing Object Detection

Ensemble methods are a reliable way to combine several models achieve superior performance. However, research on the application of ensemble in remote sensing object detection scenario is mostly overlooked. Two problems arise. First, one unique characteristic Oriented Bounding Boxes (OBB) objects and fusion multiple OBBs requires further attention. Second, widely used deep learning detectors provide score for each detected as an indicator confidence, but how use these indicators effectively method remains problem. Trying address problems, this paper proposes OBBStacking, that compatible with combines results learned fashion. This helps take 1st place Challenge Track Fine-grained Object Recognition High-Resolution Optical Images , which was featured xmlns:xlink="http://www.w3.org/1999/xlink">2021 Gaofen Automated Earth Observation Image Interpretation . The experiments DOTA dataset FAIR1M demonstrate improved performance OBBStacking features analyzed. Code will be available at https://github.com/Haoning724/obbstacking