[Re] CLRNet: Cross Layer Refinement Network for Lane Detection

Viswesh N,Kaushal Jadhav,Avi Amalanshu,Bratin Mondal,Sabaris Waran,Om Sadhwani,Apoorv Kumar,Debashish Chakravarty

2023-10-02

Abstract:The following work is a reproducibility report for CLRNet: Cross Layer Refinement Network for Lane Detection. The basic code was made available by the author. The paper proposes a novel Cross Layer Refinement Network to utilize both high and low level features for lane detection. The authors assert that the proposed technique sets the new state-of-the-art on three lane-detection benchmarks

Computer Vision and Pattern Recognition,Robotics

What problem does this paper attempt to address?

The paper attempts to address several key issues in lane detection:

1. **High Accuracy and Robustness**: The paper proposes a new network architecture—Cross Layer Refinement Network (CLRNet), which aims to improve the accuracy and robustness of lane detection by leveraging both high-level and low-level features. Traditional lane detection methods often rely on parametric techniques and predefined features such as anchors and segmentation maps. In contrast, CLRNet achieves more precise lane localization by classifying high-level and low-level features.

2. **Utilization of Global Context Information**: Under some extreme lighting conditions, lane markings may be unclear or obscured. To tackle this challenge, the paper introduces the ROIGather module, which can establish the relationship between the region of interest (ROI) lane features and the entire feature map, thereby capturing important global context information.

3. **New Loss Function**: The paper defines a new evaluation metric and loss function—Line Intersection over Union (LineIoU) loss. Unlike the traditional Intersection over Union (IoU) metric, LineIoU considers the regression of the entire lane rather than just specific points. The authors claim that training the network with this loss function can significantly improve performance, especially in complex scenarios.

4. **Performance on Multiple Datasets**: The paper claims that its method achieves state-of-the-art results on three major lane detection benchmark datasets (CULane, TuSimple, and LLAMAS). Experiments conducted on these datasets validate the effectiveness and generalization capability of the method.

In summary, the main goal of the paper is to address common issues of insufficient accuracy and poor robustness in lane detection by proposing a new network architecture and loss function, and to validate its effectiveness across multiple datasets.