Dense Pseudo-Labels based Semi-supervised Object Detection for Remote Sensing⋆

Deep-learning-based object detection has recently played an increasingly important role in analyzing geographic spatial information. However, object detection performance is strongly correlated with the quality and quantity of manually labeled data. Furthermore, object detection in Remote Sensing differs from natural scenes and faces two challenges: 1) Dense instance distribution and 2) Significant scale variations. These issues also contribute to the difficulty of manual annotation. To this end, based on a dense pseudo-labeling framework and multi-scale learning, this paper proposes a novel semi-supervised object detection (SSOD) framework for remote sensing, namely RemoteDPL. Firstly, a fusion module is proposed to adaptively integrate spatial and channel features of images at different scales, improving the detection of objects at various scales. Second, a specialized branch predicts instance density and aids in pseudo-label mining to enhance detection in dense scenarios. Finally, a two-stage filtering strategy is devised for pseudo-label mining, which first filters to obtain the "pending" class prediction boxes and then further filters this portion of prediction boxes to obtain the pseudo-labels according to the joint confidence based on the classification and density scores. Extensive experiments on DOTA-v1.0 have demonstrated that our proposed RemoteDPL surpasses the current state-of-the-art SSOD methods in various semi-supervised settings.