HG-Lane: High-Fidelity Generation of Lane Scenes under Adverse Weather and Lighting Conditions without Re-annotation
New framework boosts lane detection AI performance by up to 38.8% in snow and fog.
A research team led by Daichao Zhao has introduced HG-Lane, a novel AI framework accepted by CVPR 2026 that solves a critical data bottleneck in autonomous driving. Current lane detection models are trained on datasets like CULane and TuSimple, which lack sufficient examples of extreme weather and poor lighting, leading to unreliable and potentially dangerous performance. HG-Lane addresses this by synthetically generating high-fidelity lane scenes under conditions like heavy rain, snow, fog, and darkness. Crucially, it does this without requiring engineers to manually re-annotate the lanes in each new synthetic image, a process that is traditionally slow and expensive.
The framework was used to construct a comprehensive new benchmark containing 30,000 diverse adverse-condition images. When used to augment training data, HG-Lane delivered dramatic performance gains for state-of-the-art lane detection networks. Testing with the CLRNet model showed an overall performance (mF1 score) increase of 20.87%. The improvements were even more pronounced in specific challenging categories, with F1 scores jumping by 38.8% for snow scenes, 26.84% for fog, and 21.5% for night driving. The code and dataset have been made publicly available, providing a powerful new tool for the computer vision and automotive AI communities to build more robust and safer perception systems.
- Generates synthetic training data for snow, rain, fog, and night driving without manual lane re-annotation.
- Created a public benchmark of 30,000 high-fidelity adverse-weather lane scene images.
- Boosted CLRNet lane detection performance by 20.87% overall and by 38.8% specifically in snow conditions.
Why It Matters
Enables safer self-driving cars by providing massive, cheap training data for rare but critical dangerous road conditions.