Splitting detectors: semisupervised remote sensing object detection

Our work revolves around remote sensing object detection, where the scarcity of annotation information poses a significant challenge and hinders adequate training. To address this issue, we concentrate on semi-supervised object detection methods (SSOD), which offer a promising solution into alleviate the problem of limited labeling information.However, in our specific scenario, the self-training nature of these methods, coupled with the filtering mechanism of pseudo-labeled frames, tends to amplify noise due to the arbitrary angular rotation of remote sensing data objects. In light of these challenges, we propose a novel remote sensing task-specific SSOD framework called Splitting Detectors (SLD), which mitigates noise accumulation caused by remote sensing data characteristics. SLD has the following two innovations: (1) The SLD decouples detectors, mitigating task conflicts and thus reducing noise accumulation. (2) The SLD uses the mean-teacher semi-supervised method to train the classification head and the self-supervised method for the regression head respectively, which reduce inaccurate pseudo coordinates error caused by rotation. The Experiments on dota-split datasets demonstrate the considerable superiority of our proposed framework to other state-of-the-arts.