Jointly Enhancing Representation and Supervision for Semisupervised Rotated Object Detection of Remote Sensing Images

This is the authors' abstract. We don't add key points for this paper.

Although semisupervised learning alleviates dependence on extensive labeled data, the paradigm remains suboptimal for rotated object detection due to the following limitations. First, the limited labeled data fails to provide sufficient supervision for rotated representation learning. Second, discrete pseudolabels only preserve sparse label-level supervision, discarding continuous feature-level rotation constraints. To address the problems above-mentioned, a novel semisupervised rotated object detection (SROD) model is proposed, which jointly considers dynamic feature modulation and dense feature complement to enhance rotation-sensitive representation learning and continuous rotation constraint modeling. Specifically, a feature modulation mechanism is proposed to modulate feature aggregation according to spatial rotation characteristics, thereby boosting rotation-sensitive representation learning for diverse rotated objects. Furthermore, a feature complement mechanism is proposed to introduce continuous feature-level rotation constraints, effectively complementing the discrete supervision and guiding the model to capture accurate rotation cues. Extensive experiments are conducted on the DOTA and DIOR-R datasets, and the results show that SROD improves mean average precision by 1.25 and 0.97 points, respectively, outperforming existing semisupervised rotated object detectors.