A Comparative Study of 2D Lane Detection Paradigms: From Pixel-Level Segmentation to Structured Representations

Lane detection is a foundational task in autonomous driving, providing critical structural information for downstream modules such as trajectory prediction and motion planning. In recent years, deep learning-based 2D lane detection methods have evolved from pixel-wise segmentation toward structured representations. To systematically analyze the performance and characteristics of different paradigms, this study investigates four representative approaches: the segmentation-based RESA, the keypoint-based SRLane, the parametric curve regression-based BezierLaneNet, and the detection-based CLRerNet. The experiments were conducted on the VIL-100 dataset with high-quality annotations and diverse road scenarios. The evaluation framework covered cross-dataset fine-tuning and fully supervised training, systematically assessing the adaptability, stability, and efficiency of each method. Based on experimental results, CLRerNet achieves the best overall accuracy, while BezierLaneNet demonstrates superior computational efficiency. SRLane performs exceptionally well in curve scenes, whereas RESA consistently ranks lowest in both accuracy and efficiency. CLRerNet maintains advantages across most scenarios, though all methods struggle with complex conditions. The results indicate that structured methods more effectively leverage pre-training knowledge compared to segmentation approaches.