Abstract

We propose SC-Lane, a novel framework for 3D lane detection that introduces precise height estimation through slope-aware modeling and temporal consistency. Our approach addresses the critical challenge of accurate 3D lane reconstruction by developing a more refined height estimation framework that considers road slope variations and maintains consistency across temporal frames. Additionally, we introduce the first comprehensive height estimation benchmark specifically designed for lane detection tasks, providing standardized evaluation metrics for the research community. SC-Lane demonstrates superior performance in 3D lane detection accuracy and robustness, particularly in challenging scenarios with varying road conditions and complex geometric structures.

SC-Lane introduces a Slope-Aware Adaptive Feature module that dynamically predicts fusion weights from image cues for integrating multi-slope representations into a unified heightmap, improving robustness to diverse road geometries. A Height Consistency Module enforces temporal coherence across consecutive frames, crucial for real-world driving scenarios. We employ three standardized metrics (MAE, RMSE, and threshold-based accuracy) for road height assessment and benchmark our method using LiDAR-derived heightmap data. Extensive experiments on OpenLane demonstrate SC-Lane's state-of-the-art performance with an F-score of 64.3%, significantly outperforming existing methods.

SC-Lane Framework

Results

Heightmap Consistency Demonstration

The GIF demonstrates how consistently our framework predicts heightmaps across consecutive frames, showing the temporal stability of our height estimation approach.

3D Lane Detection Performance Comparison

The table shows that using our consistent heightmap approach, SC-Lane achieves state-of-the-art performance in 3D lane detection, outperforming all existing methods with an F-score of 64.3%.

Citation

Acknowledgements

We thank the research community for providing the benchmarks and datasets used in this research. This work was supported by ME & IPAI at Seoul National University.
The website template was borrowed from Michaël Gharbi.