Vehicle localisation and counting from high-altitude imagery are essential yet challenging tasks in remote sensing and traffic monitoring, particularly in cross-scene scenarios where the target domain differs significantly from the training data. This is especially critical for methods based on Deep Learning models, which, despite their strong performance, typically require large amounts of annotated and diverse training data to generalise effectively. In this work, we investigate the robustness and generalisation capability of a single state-of-the-art CNN-based method for vehicle localisation, evaluating its cross-scene performance across four publicly available aerial vehicle datasets. To further assess the method’s behaviour in challenging real-world conditions, we design additional experiments using a deliberately imbalanced test set composed of rare and atypical scenes—corresponding to the tails of the data distribution—such as extreme vehicle densities or uncommon layouts. Our results provide insight into the limitations of current approaches in generalising to novel domains and rare cases, highlighting important considerations for deploying vehicle localisation and counting systems in the wild.