LaneCorrect: Self-supervised Lane Detection

TL;DR

This work tackles lane detection without human labels by leveraging unsupervised 3D lane segmentation from LiDAR intensity to generate noisy 2D lane cues. A self-supervised lane correction (LaneCorrect) network enforces geometric consistency and instance-aware representation through online/target branches and contrastive learning, refining pseudo labels without ground-truth annotations. The corrected labels are then distilled to train a student detector on a target domain, enabling end-to-end inference without LiDAR at test time. Across TuSimple, CULane, CurveLanes, and LLAMAS, LaneCorrect achieves competitive results with supervised methods and shows notably better cross-domain generalization, demonstrating strong scalability and practical impact for autonomous driving deployments.

Abstract

Lane detection has evolved highly functional autonomous driving system to understand driving scenes even under complex environments. In this paper, we work towards developing a generalized computer vision system able to detect lanes without using any annotation. We make the following contributions: (i) We illustrate how to perform unsupervised 3D lane segmentation by leveraging the distinctive intensity of lanes on the LiDAR point cloud frames, and then obtain the noisy lane labels in the 2D plane by projecting the 3D points; (ii) We propose a novel self-supervised training scheme, dubbed LaneCorrect, that automatically corrects the lane label by learning geometric consistency and instance awareness from the adversarial augmentations; (iii) With the self-supervised pre-trained model, we distill to train a student network for arbitrary target lane (e.g., TuSimple) detection without any human labels; (iv) We thoroughly evaluate our self-supervised method on four major lane detection benchmarks (including TuSimple, CULane, CurveLanes and LLAMAS) and demonstrate excellent performance compared with existing supervised counterpart, whilst showing more effective results on alleviating the domain gap, i.e., training on CULane and test on TuSimple.

Figures (8)

• Figure 1: Comparison between our self-supervised lane detection method and the supervised alternative. (a) is the supervised approach, which relies on the human annotations as supervision. (b) is our LaneCorrect, which leverages point clouds to generate noisy lane clues at first and trains a lane correction network in the self-supervised manner. No human annotations are introduced in our approach.
• Figure 2: Best viewed in color and lane instance number. Our LaneCorrect consists of two collaborative networks, namely online SLC (updated by gradient descend) and target SLC (updated by moving average), to consistently correct the noisy annotation. During training, two different augmented views of pseudo lane annotations are concatenated with images and fed into online and target branches, of which outputs are collected for consistency regularization and instance similarity learning. In consistency regularization, predicted lanes of two branches are constrained to map to the unique noise-free lane locations. In instance similarity learning, multi-objective contrastive learning is adopted to ensure superior lane representation ability. During testing, only online SLC is used to predict refined lanes.
• Figure 3: The pipeline of pseudo label refinement and distillation. To enable our SLC to end-to-end inference and better align downstream datasets, we propose a pseudo-label refinement approach in the form of distillation.
• Figure 4: Incorrect annotated cases in LLAMAS. Our method can generate correct predictions in these incorrect labeled scenarios.
• Figure 5: Visualization of LaneCorrect method on multiple benchmarks compared with supervised counterpart. The top row is performance on TuSimple and the bottom row is performance on LLAMAS. The rest middle rows are qualitative results on CULane. For each row, from left to right are: input image, ground truth, results of supervised counterpart and our LaneCorrect.