Retraction Notice: This paper was formally retracted by the authors following its acceptance to AAAI 2025. The voluntary retraction was due to the accidental omission of a comparative appendix detailing the work's relationship to the authors' previous publications. This summary reflects the technical contributions described in the original manuscript.
Introduction: The Challenge of Lane Detection
Lane detection is a cornerstone of autonomous driving systems, providing the spatial boundaries necessary for safe navigation. However, real-world environments present significant "edge case" challenges that standard models often fail to navigate. These include sharp curves, heavy glare, occlusions from other vehicles, and sections where lane markings are degraded or entirely absent.
The authors of GEONet argue that existing methods struggle because they rely too heavily on local features or lightweight backbones. Without a holistic "global" understanding of the road scene, models cannot effectively infer the position of a lane when visual evidence is weak. GEONet (Global Enhancement and Optimization Network) was designed to bridge this gap by prioritizing global feature transmission and refinement.
Methodology: Enhancing Global Context
The GEONet architecture introduces several novel components to capture and refine global information:
1. Global Feature Extraction Module (GFEM)
The GFEM utilizes a dual-branch structure to balance different types of global information. One branch employs an MLP-mixer to capture coarse, foundational global features across the entire image. The second branch uses dilated convolutions to expand the receptive field, capturing fine-grained details that might be missed by a purely global approach.
2. Top-Tier Supplementary Module (TTSM)
Positioned at the highest level of the network, the TTSM uses self-attention mechanisms to distill features. Rather than just extracting more data, it identifies and refines the most relevant global semantic information, ensuring that the final decision-making layers have a clear "understanding" of the road layout.
3. Whitening Modules (WBN & WCL)
To improve robustness, GEONet replaces standard Batch Normalization with Whitening Batch Normalization (WBN). This process decorrelates features, reducing redundancy and making the network more resilient to domain shifts and complex lighting. Additionally, Whitening Contrastive Learning (WCL) is used during pre-training to create a more discriminative embedding space.
Geometric Optimization
Standard loss functions often fail to account for the thin, elongated geometry of lanes. GEONet introduces two specialized geometric loss functions:
- GRIoU (Generalized Rectangular IoU): This loss treats lane segments as rectangles rather than points, allowing for better optimization of overlap and spatial smoothness.
- Angle Loss: This function specifically aligns the orientation of predicted lanes with the ground truth, which significantly improves continuity in curved sections.
Experimental Results
GEONet was evaluated on the CULane and TuSimple benchmarks, achieving state-of-the-art (SOTA) performance:
- CULane: The model achieved an F1@50 score of 82.18 using a DLA34 backbone. It outperformed previous methods particularly in challenging categories like "Cross" and "Night."
- TuSimple: GEONet set a new benchmark with an F1 score of 98.21, while maintaining a very low false-positive rate of 1.83%.
Ablation studies confirmed that the GFEM and TTSM modules were critical to these gains, particularly in "No line" and "Shadow" scenarios where global context is most necessary.
Conclusion
GEONet demonstrates that prioritizing global context and geometric consistency can significantly improve lane detection in difficult environments. While the paper was retracted due to administrative omissions regarding its relationship to prior work, the technical framework—specifically the use of feature whitening and global distillation—offers valuable insights for the field of autonomous perception.
Summary updated