Revisiting Vehicle Color Recognition in Long-Tailed Surveillance Scenarios

A new study revisits vehicle color recognition for surveillance, reporting a macro accuracy of 79.7 percent on the challenging UFPR-VeSV dataset by combining synthetic data generation with ensemble fusion techniques, according to a paper submitted June 11, 2026. [1] Vehicle color serves as a critical identifier in surveillance systems when license plates are obscured by low resolution, motion blur, or poor lighting. [1] The research addresses a persistent problem: real-world color distributions are heavily skewed toward common shades, meaning standard accuracy metrics can mask poor performance on rare but operationally important colors. [1] The team investigated two off-the-shelf generative strategies to augment minority classes. One approach used text-conditioned image generation with RunDiffusion/JuggernautXL, while the other employed image-conditioned color editing with Gemini 2.0 Flash. [1] These synthetic samples were then integrated into a pipeline that included modern visual representations, loss reweighting, learning-rate scheduling, color-safe augmentation, foreground-aware preprocessing, and ensemble fusion. [1] The best-performing configuration achieved 94.6 percent micro accuracy and 79.7 percent macro accuracy, representing an improvement of 8.2 percentage points in macro accuracy over recent literature. [1] The generated images and source code have been made publicly available. [1] A manual error analysis revealed that many of the remaining classification failures involved images that were visually ambiguous even for human annotators. [1] This finding underscores the inherent limitations of relying solely on color for vehicle identification in unconstrained surveillance imagery. [1]