Unsupervised Vehicle Re-Identification with Clustering and Appearance Style Exploration

Vehicle re-identification (Re-ID) has emerged as a critical challenge in the fields of computer vision and intelligent transportation systems, driven by the need to accurately identify vehicles across multiple camera views in diverse urban environments. Traditional methods relying on appearance-based features such as color and resemblance often fall short due to the variability in vehicle appearances caused by factors like illumination changes, camera angles, and occlusions. To address these limitations, this paper proposes a novel approach that leverages unsupervised learning, integrating agglomerative clustering techniques and style transfer via the StarGAN2 model. By treating each camera as a unique style domain, the proposed method generates new training pictures with diverse styles, reducing camera-induced variability and enhancing cross-camera vehicle identification. The repeated learning framework further refines vehicle identities by clustering similar pictures, thereby improving the system's ability to distinguish between vehicles in real-world scenarios. This approach not only minimizes the need for manual annotation but also enhances the robustness and adaptability of vehicle Re-ID systems, making it a valuable contribution to urban surveillance and intelligent transportation systems.