Few-Shot Object Detection in Remote Sensing: Mitigating Label Inconsistencies and Navigating Category Variations

Abstract

Over recent years, the increasing expansion of remote sensing image (RSI) datasets has made annotation tasks more challenging and labor-intensive, drawing considerable attention toward few-shot object detection (FSOD). Nevertheless, current mainstream FSOD models are primarily designed for natural images and encounter two substantial challenges when applied to RSIs. 1) Inconsistent label assignment for novel instances between pre-training and fine-tuning confuses detectors, leading to diminished generalization performance. 2) Complex scenes within RSIs result in significant category variations, comprising high inter-class similarity and large intra-class variance, which impairs classification accuracy. Against the aforementioned challenges, we propose a novel FSOD approach in RSIs, termed EC-FSOD. Specifically, our approach introduces two key modules: Ensemble Class-free RPN (ECF-RPN) and Contrastive Prototype ETF Classifier (CPEC). The preceding module, ECF-RPN, generates proposals by integrating multiple dissimilar yet cooperative Class-free RPNs that perceive the shape and location of target objects, mitigating the confusion caused by label inconsistencies. Furthermore, the subsequent CPEC module combines two submodules, namely Contrastive Prototype Learning Network (CPLN) and Simplex ETF Classifier (SEC), to obtain a set of representative class prototypes and robust discriminative feature representations, which are employed to overcome the category variations and enhance the generalization performance of novel instances. Extensive experiments have revealed that our approach achieves top-2 results on the DIOR dataset and optimal performance on the NWPU VHR-10.v2 dataset.

Keywords

• Few-shot object detection (FSOD)
• remote sensing images (RSIs)
• transfer learning
• metric learning