CLRerNet: Improving Confidence of Lane Detection With LaneIoU

Hiroto Honda, Yusuke UchidaWACV 2024

TL;DR

This paper introduces LaneIoU, a novel metric that better correlates with the standard segmentation-based IoU for lane detection by considering local lane angles. By integrating LaneIoU into an anchor-based lane detection framework, they developed CLRerNet, which achieves state-of-the-art performance on challenging lane detection benchmarks.

What Problem Does This Paper Address?

Accurate lane detection is crucial for autonomous driving and driver assistance systems. While modern deep learning methods, particularly anchor-based approaches, have achieved impressive results, there's a persistent gap between predicted lane confidence scores and the actual evaluation metric (segmentation-based Intersection over Union or IoU). This disconnect means that even when a lane is correctly detected, its confidence score might not accurately reflect its quality, leading to suboptimal performance. The paper hypothesizes that the core issue lies in how lane similarity is measured during training, especially for non-vertical or curved lanes, where standard metrics fail to capture the true overlap with ground truth.

Methods

The authors build upon the CLRNet anchor-based lane detection framework. Their primary contribution is the LaneIoU metric. Unlike previous methods that used horizontal distance or a simplified IoU that didn't account for lane curvature, LaneIoU calculates overlap and union by considering the local tilt of lane segments at each row. This makes it more robust to variations in lane angles and better aligned with the segmentation-based IoU used for evaluation.

LaneIoU is integrated into the CLRerNet framework in three key ways:

  1. Sample Assignment Cost: LaneIoU is used to dynamically assign anchors to ground truth lanes, prioritizing those with higher LaneIoU. This ensures that anchors with better potential overlap are selected for training.
  2. Assignment Cost Function: LaneIoU is incorporated into the cost matrix for sample assignment, further refining the selection of positive anchors.
  3. Loss Function: LaneIoU replaces the previous LineIoU loss, providing a more accurate penalty for incorrect lane predictions, especially for tilted or curved lanes.

The paper also emphasizes the importance of a fair benchmarking protocol, including multi-model evaluation (averaging results from five models trained with different random seeds) and 5-fold cross-validation for determining optimal confidence thresholds.

Datasets

The paper evaluates CLRerNet on two widely used lane detection datasets:

  • CULane: A standard benchmark dataset with 88,880 training frames, featuring various scene types including normal, crowded, and dazzling conditions.
  • CurveLanes: A dataset specifically designed to challenge models with complex curved lane scenarios.

Key Insights and Ideas

  • Confidence is King (When Measured Right): Oracle experiments revealed that existing detectors already predict correct lane positions with high probability. The key to improving performance lies in accurately predicting the confidence of these detections, which should directly correlate with the evaluation metric (IoU).
  • Angles Matter for Lane IoU: Standard IoU calculations, especially those that don't account for local lane orientation, can be misleading for tilted or curved lanes. LaneIoU's angle-aware approach bridges this gap, leading to better correlation with ground truth IoU.
  • Fair Benchmarking is Crucial: The paper highlights the sensitivity of F1 scores to confidence thresholds and the variability introduced by random seeds. Their rigorous benchmarking protocol (multi-model averaging and cross-validation) provides a more reliable and fair comparison of methods.
  • Anchor Assignment Benefits from LaneIoU: Dynamically assigning anchors based on LaneIoU leads to more homogeneous anchor assignments across different lane angles, preventing suboptimal training due to underestimating IoU for non-vertical lanes.

Key Results

  • State-of-the-Art Performance: CLRerNet achieves state-of-the-art results on both CULane and CurveLanes.
  • On CULane, CLRerNet (DLA34 backbone) achieves an F1 score of 81.43%, surpassing the previous best of 80.47% (CLRNet).
  • On CurveLanes, CLRerNet achieves 86.47%, a 0.37% improvement over the baseline.
  • Significant Improvement from LaneIoU: Even without architectural changes, integrating LaneIoU into existing models like CLRNet and LaneATT leads to noticeable performance gains (e.g., +0.26% F150 for CLRNet-DLA34).
  • No Increase in Computational Cost: CLRerNet maintains the same GFLOPs and FPS as its baseline CLRNet, meaning the performance gains come without a computational penalty.
  • Improved Confidence Accuracy: Qualitative analysis (Figure 7) shows that CLRerNet has a lower error between predicted confidence scores and metric IoU, particularly in non-vertical angle ranges.

Limitations and Future Work

  • Dataset-Oriented Issues: The authors acknowledge that dataset-specific challenges like label fluctuation and data imbalance (especially in "Noline" scenes) contribute to the remaining performance gap. Mitigating these issues is suggested as a future direction.
  • Oracle Gap: Despite achieving state-of-the-art, there's still a significant gap between CLRerNet's performance (81.43%) and the oracle-confidence case (98.47%), indicating room for further improvement.
  • Generalizability: While LaneIoU shows promise for anchor-based methods, its applicability to other lane detection paradigms (e.g., segmentation-based) is not explored.

Summary updated

BibTeX
@inproceedings{honda2024clrernet,
  author = {Hiroto Honda and Yusuke Uchida},
  title = {{CLRerNet: Improving Confidence of Lane Detection With LaneIoU}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure