[Paper Review] CLRerNet: Improving Confidence of Lane Detection with LaneIoU

Hiroto Honda, Yusuke Uchida|arXiv (Cornell University)|May 15, 2023

Autonomous Vehicle Technology and Safety4 citations

TL;DR

CLRerNet improves lane detection confidence by introducing LaneIoU, a novel IoU metric that accounts for local lane angles to better align predicted confidence scores with segmentation-based IoU. By integrating LaneIoU into sample assignment and loss functions in a row-based detector, CLRerNet achieves state-of-the-art F1 scores of 81.43% on CULane and 86.47% on CurveLanes, significantly outperforming prior methods through a fair, five-fold cross-validated benchmark.

ABSTRACT

Lane marker detection is a crucial component of the autonomous driving and driver assistance systems. Modern deep lane detection methods with row-based lane representation exhibit excellent performance on lane detection benchmarks. Through preliminary oracle experiments, we firstly disentangle the lane representation components to determine the direction of our approach. We show that correct lane positions are already among the predictions of an existing row-based detector, and the confidence scores that accurately represent intersection-over-union (IoU) with ground truths are the most beneficial. Based on the finding, we propose LaneIoU that better correlates with the metric, by taking the local lane angles into consideration. We develop a novel detector coined CLRerNet featuring LaneIoU for the target assignment cost and loss functions aiming at the improved quality of confidence scores. Through careful and fair benchmark including cross validation, we demonstrate that CLRerNet outperforms the state-of-the-art by a large margin - enjoying F1 score of 81.43% compared with 80.47% of the existing method on CULane, and 86.47% compared with 86.10% on CurveLanes.

Motivation & Objective

• To identify the key bottleneck in lane detection performance, particularly the misalignment between predicted confidence scores and the actual IoU metric used in evaluation.
• To address the poor correlation between standard confidence scores and segmentation-based IoU, especially for tilted or curved lanes.
• To develop a training method that directly optimizes confidence scores to reflect true IoU with ground truth lanes.
• To establish a fair and reliable benchmarking protocol using cross-validation and ensemble model averaging to ensure reproducibility and robustness.

Proposed method

• Propose LaneIoU, a novel intersection-over-union metric that incorporates local lane angles to better approximate segmentation-based IoU for non-vertical lanes.
• Integrate LaneIoU into the dynamic sample assignment process to prioritize anchors with higher metric IoU, improving confidence score learning.
• Replace the standard LineIoU loss with LaneIoU loss to reduce tilt-dependent regression bias and improve localization accuracy.
• Develop CLRerNet, a new detector based on CLRNet, by replacing the confidence prediction head and loss functions with LaneIoU-aware components.
• Use a five-seed, five-fold cross-validation protocol to compute average performance, ensuring fair and robust evaluation.
• Apply the empirically optimal confidence threshold (0.44) via cross-validation to maximize F1 score on validation splits.

Experimental results

Research questions

• RQ1 Why do state-of-the-art lane detectors underperform despite detecting correct lane positions?
• RQ2 Can confidence scores be improved to better reflect the true IoU with ground truth lanes, especially for tilted or curved lanes?
• RQ3 Does incorporating local lane angle information into the IoU computation lead to more accurate and consistent confidence predictions?
• RQ4 How does using a geometry-aware IoU metric (LaneIoU) in sample assignment and loss functions affect detector performance?
• RQ5 What is the impact of a rigorous, cross-validated benchmarking protocol on the reliability of reported performance gains?

Key findings

• CLRerNet achieves a new state-of-the-art F1 score of 81.43% on the CULane dataset, outperforming the previous SoTA method by 0.96 percentage points.
• On CurveLanes, CLRerNet achieves an F1 score of 86.47%, surpassing the prior SoTA by 0.37 percentage points.
• The ablation study confirms that replacing LaneIoU with LineIoU in dynamic-k assignment, cost function, and loss function leads to consistent performance degradation across all settings.
• The $l1$ error between predicted confidence scores and ground-truth IoU is significantly reduced in non-vertical angle ranges, demonstrating improved confidence calibration.
• The number of assigned anchors per GT becomes more homogeneous across different lane angles when using LaneIoU, reducing bias from tilt variation.
• The oracle experiment reveals that confidence scores are the primary bottleneck—when set to true IoU values, F1 reaches 98.47%, indicating that current models fail to predict confidence accurately despite detecting correct positions.