A COMPREHENSIVE REVIEW AND EXPERIMENTAL ANALYSIS OF CAR PART DETECTION AND INSTANCE SEGMENTATION FOR AUTOMOTIVE APPLICATIONS

Abstract

Car-part detection and instance segmentation play a critical role in modern automotive applications, including automated vehicle inspection, repair assistance, insurance assessment, and intelligent transportation systems. Unlike generic vehicle detection, fine-grained recognition of individual car components requires specialized datasets and efficient algorithms capable of operating in real-time. This paper presents a comprehensive review and experimental analysis of recent advances in car part detection and segmentation. Publicly available datasets of varying scale and annotation strategies are examined, and the performance of representative two-stage and one-stage deep-learning models is analyzed using reported experimental results. Particular attention is given to lightweight YOLO-based segmentation models and large-scale datasets that enable practical deployment. The results demonstrate that compact one-stage architectures can achieve competitive accuracy when trained on sufficiently large and diverse datasets, while offering significant advantages in inference speed and computational efficiency. Furthermore, semi-supervised annotation strategies combining manual and automatically generated labels are shown to be effective in reducing labeling cost without compromising performance. The findings highlight key trade-offs between accuracy, efficiency, and scalability and provide insights for the development of robust car part detection systems in real-world automotive environments.