The field of computer vision and remote sensing is moving towards label-efficient learning, with a focus on developing methods that can learn effectively from limited or partially annotated data. This direction is driven by the need to leverage abundant unlabeled data in real-world applications, such as Earth observation and object detection. Recent advances in weakly supervised learning, self-supervised learning, and few-shot learning have shown promising results in improving the performance of models in these areas. Notably, the use of heterogeneous network architectures, dual spectral enhancement techniques, and adversarial class prompting have been shown to enhance the robustness and accuracy of models. Furthermore, the development of large-scale datasets, such as IRSAMap, and the application of semi-supervised and annotation-free methods have expanded the possibilities for label-efficient learning in remote sensing.

Some noteworthy papers in this area include: Through the Looking Glass, which proposes a novel heterogeneous network architecture for weakly-supervised few-shot segmentation, achieving a 13.2% improvement on Pascal-5i and a 9.7% improvement on COCO-20i. IRSAMap, which introduces a large-scale dataset for land cover vector mapping, providing a comprehensive vector annotation system and an intelligent annotation workflow. Advancing Weakly-Supervised Change Detection in Satellite Images via Adversarial Class Prompting, which proposes an Adversarial Class Prompting method to address the co-occurring noise problem in weakly-supervised change detection, demonstrating significant performance enhancements on various baselines.