CLRmatchNet: Enhancing Curved Lane Detection with Deep Matching Process

TL;DR

The paper addresses the problem that traditional label assignment in lane-detection models underperforms for curved lanes due to fixed, handcrafted cost functions. It proposes CLRmatchNet, which replaces the classic label assignment with MatchNet, a neural scoring module that learns to assign predictions to ground-truth lanes, integrated into CLRNet using a teacher-student training scheme and a dynamic k selection. The approach improves curved-lane detection across multiple backbones, raises prediction confidence, and demonstrates that deep-label assignment can outperform fixed-cost strategies while maintaining performance on other scenarios; the model is trained with a joint loss $\mathcal{L} = \lambda_{0}\mathcal{L}_{xyl\theta} + \lambda_{1}\mathcal{L}_{cls} + \lambda_{2}\mathcal{L}_{seg} + \lambda_{3}\mathcal{L}_{LineIoU}$ and uses dynamic matching up to $k=4$ per GT. The work has practical significance for autonomous driving by improving robustness to curved lanes and increasing the reliability of lane detections, with code released at the provided repository.

Abstract

Lane detection plays a crucial role in autonomous driving by providing vital data to ensure safe navigation. Modern algorithms rely on anchor-based detectors, which are then followed by a label-assignment process to categorize training detections as positive or negative instances based on learned geometric attributes. Accurate label assignment has great impact on the model performance, that is usually relying on a pre-defined classical cost function evaluating GT-prediction alignment. However, classical label assignment methods face limitations due to their reliance on predefined cost functions derived from low-dimensional models, potentially impacting their optimality. Our research introduces MatchNet, a deep learning submodule-based approach aimed at improving the label assignment process. Integrated into a state-of-the-art lane detection network such as the Cross Layer Refinement Network for Lane Detection (CLRNet), MatchNet replaces the conventional label assignment process with a submodule network. The integrated model, CLRmatchNet, surpasses CLRNet, showing substantial improvements in scenarios involving curved lanes, with remarkable improvement across all backbones of +2.8% for ResNet34, +2.3% for ResNet101, and +2.96% for DLA34. In addition, it maintains or even improves comparable results in other sections. Our method boosts the confidence level in lane detection, allowing an increase in the confidence threshold. Our code is available at: https://github.com/sapirkontente/CLRmatchNet.git

Figures (9)

• Figure 1: Visualizing lane detection results: a comparison between CLRNet and CLRmatchNet (our approach) of CULane testing set in the curve category
• Figure 2: CLRmatchNet architecture: enhancing CLRNet with MatchNet integration. Gray blocks represent CLRNet's original components, and the pink block depicts MatchNet, which replaces the classic cost function, indicated as "deleted".
• Figure 3: Architecture of MatchNet. A fully connected model for match classification. As an input, MatchNet receives geometrical parameters of $N$ predictions in relative to each one of the GT, and outputs probabilities defining whether the each pair is positive or negative.
• Figure 4: MatchNet teacher-student training setup. The gray blocks represent the teacher model (CLRNet) and the lower blue blocks represent the student model (MatchNet). In this setup, for each image, the teacher predicts N lanes and computes the cost parameters between each prediction i and each GT j in the image. These parameters are used as input to MatchNet. To determine if the pair (i, j) is a match, we utilize both the classical cost function decision and a feedback from the loss. Only when both criteria are met (set to 1), we designate this pair as a match for reference. MatchNet is trained using the inputs and references supplied by the CLRNet.
• Figure 5: Visualization results of CLRNet and our method for night, shadow, and crowd scenarios of the CULane dataset (from top to bottom)