Weakly Semi-Supervised Oriented with Points for Remote Sensing Vehicle Detection

Vehicle detection is crucial for urban planning and traffic management, with remote sensing imagery serving as a major vehicle data source. Vehicles in remote sensing images are characterized by dense distribution and arbitrary orientations. Traditional object detection tasks with horizontal bounding box annotations often include excessive background information, leading to lower accuracy. Conversely, the annotation cost for oriented bounding boxes (OBB) is significantly high. To address these challenges, we propose a weakly semi-supervised vehicle detection with points. For this method, we used a teacher model to accurately translate point annotations into OBBs through instance-level proposal grouping and RoI enhancement unit. These pseudo-boxes are then used to train the student model. Our experiments on the DOTA vehicle dataset show results comparable to fully supervised models. This approach significantly mitigates the challenge of annotating OBB for dense and small-scale objects in remote sensing imagery.