ADA-IRSTD: Active Domain Adaptation for Cross-Domain Infrared Small Target Detection

Infrared small target detection (IRSTD) is critical for applications such as surveillance and counter-UAV early warning. However, existing methods often suffer from limited sample sizes and limited scene diversity, leading to domain shifts between synthetic training data and real-world test images. Although unsupervised domain adaptation (UDA) has been explored to mitigate this problem, a notable performance gap with fully supervised methods remains, limiting practical applicability. To address this, we propose a novel Active Domain Adaptation framework for IRSTD, namely ADA-IRSTD. ADA-IRSTD aims to achieve high performance at the lowest possible annotation cost by selecting the most informative samples from the target domain, thereby balancing high performance with low annotation expenses. Our ADA-IRSTD contains three simple yet effective strategies: 1) We first design Target-Proximity Preserving Masking (TPPM) to mask background regions while preserving target and neighboring features, thereby enhancing cross-domain robustness. 2) Then, we incorporate a False Alarm Prediction Module (FAPM) to quantify target indistinguishability and assess the risk of ambiguous detections. 3) Finally, we propose Global Distribution Uncertainty Sampling (GDUS) to guide active sample selection based on global prediction uncertainty. Sufficient experiments and comparisons are conducted on NUDT-SIRST, IRSTD-1 k, SIRST, and SIRST-v2 datasets. The results show that our ADA-IRSTD outperforms other active domain adaptation methods and achieves promising performance in cross-domain IRSTD tasks, e.g., in theN$\rightarrow$S$\_{1}$, ADA-IRSTD achieves 97.61% of the fully supervised method's performance with only 10% of the annotation budget.