Abstract
Lane detection plays a crucial role in autonomous driving by providing vitaldata to ensure safe navigation. Modern algorithms rely on anchor-baseddetectors, which are then followed by a label-assignment process to categorizetraining detections as positive or negative instances based on learnedgeometric attributes. Accurate label assignment has great impact on the modelperformance, that is usually relying on a pre-defined classical cost functionevaluating GT-prediction alignment. However, classical label assignment methodsface limitations due to their reliance on predefined cost functions derivedfrom low-dimensional models, potentially impacting their optimality. Ourresearch introduces MatchNet, a deep learning submodule-based approach aimed atimproving the label assignment process. Integrated into a state-of-the-art lanedetection network such as the Cross Layer Refinement Network for Lane Detection(CLRNet), MatchNet replaces the conventional label assignment process with asubmodule network. The integrated model, CLRmatchNet, surpasses CLRNet, showingsubstantial improvements in scenarios involving curved lanes, with remarkableimprovement across all backbones of +2.8% for ResNet34, +2.3% for ResNet101,and +2.96% for DLA34. In addition, it maintains or even improves comparableresults in other sections. Our method boosts the confidence level in lanedetection, allowing an increase in the confidence threshold. Our code isavailable at: https://github.com/sapirkontente/CLRmatchNet.git