Generalization-Enhanced Few-Shot Object Detection in Remote Sensing

TL;DR

GE-FSOD tackles the generalization bottleneck in remote-sensing few-shot object detection by replacing standard neck and head modules with CFPAN and MRRPN, and by adopting the Generalized Classification Loss to better handle base-to-novel class transfer. The approach simultaneously enhances multi-scale feature fusion, region proposal refinement, and robust classification through placeholder-based adaptation. Empirical results on the DIOR and NWPU VHR-10 datasets show state-of-the-art FSOD performance across 3–20 shots, with notable gains from CFPAN, CBAM, MRRPN, and GCL ablations. The work demonstrates that careful architectural and loss-function design can substantially improve few-shot generalization in high-variability remote-sensing imagery, with practical implications for rapid deployment in disaster response and urban monitoring.

Abstract

Remote sensing object detection is particularly challenging due to the high resolution, multi-scale features, and diverse ground object characteristics inherent in satellite and UAV imagery. These challenges necessitate more advanced approaches for effective object detection in such environments. While deep learning methods have achieved remarkable success in remote sensing object detection, they typically rely on large amounts of labeled data. Acquiring sufficient labeled data, particularly for novel or rare objects, is both challenging and time-consuming in remote sensing scenarios, limiting the generalization capabilities of existing models. To address these challenges, few-shot learning (FSL) has emerged as a promising approach, aiming to enable models to learn new classes from limited labeled examples. Building on this concept, few-shot object detection (FSOD) specifically targets object detection challenges in data-limited conditions. However, the generalization capability of FSOD models, particularly in remote sensing, is often constrained by the complex and diverse characteristics of the objects present in such environments. In this paper, we propose the Generalization-Enhanced Few-Shot Object Detection (GE-FSOD) model to improve the generalization capability in remote sensing FSOD tasks. Our model introduces three key innovations: the Cross-Level Fusion Pyramid Attention Network (CFPAN) for enhanced multi-scale feature representation, the Multi-Stage Refinement Region Proposal Network (MRRPN) for more accurate region proposals, and the Generalized Classification Loss (GCL) for improved classification performance in few-shot scenarios. Extensive experiments on the DIOR and NWPU VHR-10 datasets show that our model achieves state-of-the-art performance for few-shot object detection in remote sensing.

Figures (6)

• Figure 1: Overview of the proposed Generalization-Enhanced Few-Shot Object Detection (GE-FSOD) model architecture. The backbone extracts multi-scale feature maps, which are defined by the Cross-Level Fusion Pyramid Attention Network (CFPAN) to enhance multi-scale feature representation through dual attention mechanisms and cross-level feature fusion. The Multi-Stage Refinement Region Proposal Network (MRRPN) further refines the feature maps to generate accurate region proposals through a multi-stage refinement strategy. Finally, the model utilizes ROI Pooling to standardize proposals, followed by Classification and Bounding Box Regression heads for detecting and localizing objects. During the pretraining phase, all components of the model are optimized to ensure comprehensive feature learning, whereas, in the fine-tuning phase, the backbone is kept frozen while other modules are fine-tuned to effectively adapt to few-shot tasks.
• Figure 2: Illustration of the Convolutional Block Attention Module (CBAM).
• Figure 3: Visualization of FSOD results of our model on the DIOR dataset under four different splits for 3-shot, 5-shot, 10-shot, and 20-shot settings.
• Figure 4: Visualization comparison of FSOD results between our model, G-FSDet and the SAE-FSDet model on the DIOR dataset under 3-shot, 5-shot, 10-shot, and 20-shot settings.
• Figure 5: Visualization of FSOD results of our model on the NWPU VHR-10 dataset under 3-shot, 5-shot, 10-shot, and 20-shot settings.