InfRS: Incremental Few-shot Object Detection in Remote Sensing Images

Few-shot detection in remote sensing images has witnessed significant advancements recently. Despite these progresses, the capacity for continuous conceptual learning still poses a significant challenge to existing methodologies. In this paper, we explore the intricate task of incremental few-shot object detection (iFSOD) in remote sensing images. We present a pioneering transfer-learning-based technique, termed InfRS, designed to enable the incremental learning of novel classes using a restricted set of examples, while simultaneously preserving the knowledge learned from previously seen classes without the need to revisit old data. Specifically, we pretrain the detector using sufficient data from base datasets and then generate a set of class-wise prototypes that represent the intrinsic characteristics of the data. In the incremental learning session, we design a Hybrid Prototypical Contrastive (HPC) encoding module for learning discriminative representations. Furthermore, we develop a prototypical calibration strategy based on the Wasserstein distance to overcome the catastrophic forgetting problem. Comprehensive evaluations conducted with two aerial imagery datasets show that our InfRS effectively addresses the iFSOD issue in remote sensing imagery. Code is available at https://github.com/lyanna4869/InfRS.git.