SCOUT-Count: Stylization with covariance-whitening and uncertainty for domain-generalized remote sensing object counting

Object counting in remote sensing imagery faces significant challenges due to domain shift, manifested as variations in geographical background, object scale, image resolution, and imaging style across datasets. To address this, we propose SCOUT-Count, a novel domain-agnostic object counting method tailored for remote sensing scenarios. Our method is built upon a synergy of modules that jointly expand style diversity, introduce controlled perturbations, and stabilize the learning process. The Global Style Element Extraction (GSEE) module captures broad stylistic cues from the source domain, serving as a foundation for generating new variations. Complementing this, the Local Uncertainty Perturbation (LUP) module injects batch-level uncertainty-driven variations, encouraging the model to explore styles not present in the training data. To prevent instability caused by excessive stylization, the Stylization Sensitivity Covariance Whitening Loss (SS-CWL) enforces consistency between original and augmented representations, balancing fidelity and diversity throughout training. Given the specific characteristics of remote sensing data, the CLIP-based Class activation map Adjuster (CCA) module leverages the pretrained CLIP-RS model’s remote sensing semantic knowledge to improve the robustness of density maps against cluttered backgrounds. Comprehensive experiments validate the effectiveness of our approach, demonstrating its strong generalization capabilities across various datasets. Code is available at https://github.com/WindermerePeaks/SCOUT-Count.