Addressing Sample Inconsistency for Semi-Supervised Object Detection in Remote Sensing Images

The emergence of Semi-Supervised Object Detection (SSOD) techniques has greatly enhanced object detection performance. SSOD leverages a limited amount of labeled data along with a large quantity of unlabeled data. However, there exists a problem of sample inconsistency in remote sensing images, which manifests in two ways. Firstly, remote sensing images are diverse and complex. Conventional random initialization methods for labeled data are insufficient for training teacher networks to generate high-quality pseudo-labels. Secondly, remote sensing images typically exhibit a long-tailed distribution, where some categories have a significant number of instances, while others have very few. This distribution poses significant challenges during model training. In this paper, we propose the utilization of SSOD networks for remote sensing images characterized by a long-tailed distribution. To address the issue of sample inconsistency between labeled and unlabeled data, we employ a labeled data iterative selection strategy based on the active learning approach. We iteratively filter out high-value samples through the designed selection criteria. The selected samples are labeled and used as data for supervised training. This method filters out valuable labeled data, thereby improving the quality of pseudo-labels. Inspired by transfer learning, we decouple the model training into the training of the backbone and the detector. We tackle the problem of sample inconsistency in long-tail distribution data by training the detector using balanced data across categories. Our approach exhibits an approximate 1% improvement over the current state-of-the-art models on both the DOTAv1.0 and DIOR datasets.