FEATURES MATTER: CROSS-DOMAIN MUTUAL CENTRALIZED LEARNING FOR FEW SHOT REMOTE SENSING OBJECT DETECTION

We propose a novel approach for few-shot object detection by integrating weight generation and a feature memory library. To overcome the challenge of limited labeled data, our method assigns adaptive weights to the support set during training, focusing on informative samples and enabling better adaptation to unseen classes. Additionally, a feature memory library facilitates knowledge transfer by allowing the model to access and compare feature embeddings across different categories. Our approach achieves state-of-the-art results on fewshot detection benchmarks, outperforming existing methods in accuracy and robustness. Experimental results demonstrate the effectiveness of weight generation and the feature memory library in handling the few-shot detection scenario. Furthermore, our method exhibits strong generalization capabilities, adapting well to new object categories with limited labeled data. This framework provides valuable insights for remote sensing applications with limited labeled data, advancing the field of few-shot object detection.