- 1. Introduction
- 2. Contribution
- 3. Framework
- 4. Scene-Resolution Shift Benchmark
- 5. Experiments
- 6. Getting Started
The official implementation of paper "Few-Shot SAR Object Detection Under Scene–Resolution Shifts".
- We identify and formally define the Scene–Resolution Domain Shift (SR-Shift) as a critical challenge in generalizable SAR object detection, which has been largely overlooked in prior research. We introduce SR-Bench, the first benchmark to evaluate the few-shot SR-Shift problem, providing a comprehensive evaluation tool for SR-Shift-related challenges.
- We propose the Scene–Resolution Generalization Detector (SRGD), a novel framework that addresses the SR-Shift challenge by jointly handling scene and resolution variations. Unlike existing methods that treat these factors separately, SRGD decouples scene semantics from target features using a vision-language model, ensuring robust generalization with few-shot samples. Additionally, it selectively aligns low-frequency components of decoupled target features across resolutions, enabling cross-resolution generalization. This integrated approach provides a unique solution to SR-Shift, filling a gap that current methods do not address effectively.
- Experimental results demonstrate that SRGD outperforms existing methods on SR-Bench, achieving state-of-the-art performance and showcasing flexibility for both horizontal and oriented object detectors.
