Exploring key modules in lane detection through comparing two updated models

Lane detection is significant in the digital navigation system of modern vehicles. It is vita in lane switching and route planning. We want to explore what modules are critical in lane detection. In this work, we choose two representative models, Cross Layer Refinement Network (CLRNet) and Ultra Fast Lane Detector version 2 (UFLD-v2), for comparison, aiming at evaluating their performance and finding out what determines their performance or why they perform well. The first one is a top-down model while another is a bottom-top model, which represents they are built through different ideas. We use two common datasets, CULane and TuSimple, to train, validate, and test them. The experiment results are divided into quantitative results and qualitative results, by which we can conclude that global learning capability and immunity to interference are indispensable.