GAUSSIAN WASSERSTEIN METRIC BASED SEMI-SUPERVISED REMOTE SENSING OBJECT DETECTION

Remote sensing semi-supervised object detection (SSOD) endeavors to harness the unlabeled data to boost object detection performance in the case of limited labeled data, which has been primarily facilitated by the pseudo-labeling technique. However, the pseudo-label generation exhibits high uncertainties, such as unstable categories and bounding boxes across different stages of training, leading to cumulative errors in SSOD pseudo-label predictions. Furthermore, the current sample assignment based on intersection over union (IoU) cannot well reflect the quality of pseudo-labels, which are further skewed by the misassigned positive samples. To address this issue, we explore the uncertainty of pseudo-labels in remote sensing SSOD and propose a new evaluation metric that incorporates optimal transport theory into the SSOD framework, replacing the traditional IoU with the Gaussian Wasserstein Distance to more precisely evaluate the quality of pseudo-labels. Specifically, the Gaussian Wasserstein distance is applied on both sample allocation and regression loss functions, aligning the metrics for sample allocation and loss functions, which allows for optimal training of the detector. Extensive experimental results demonstrate that our method achieves a 5.8% mAP improvement over state-of-the-art (SOTA) approaches when using only 10% labeled data.