[Paper Review] Decoupling the Curve Modeling and Pavement Regression for Lane Detection

Wencheng Han, Jianbing Shen|arXiv (Cornell University)|Sep 19, 2023

Autonomous Vehicle Technology and Safety4 citations

TL;DR

This paper proposes DecoupleLane, a novel lane detection framework that decouples curve modeling in Bird's Eye View (BEV) space from ground height regression to address performance drops caused by uneven pavement. By modeling smooth lane shapes independently from road surface fluctuations, the method achieves state-of-the-art performance on both 2D and 3D lane detection benchmarks, outperforming existing curve-based and point-based approaches.

ABSTRACT

The curve-based lane representation is a popular approach in many lane detection methods, as it allows for the representation of lanes as a whole object and maximizes the use of holistic information about the lanes. However, the curves produced by these methods may not fit well with irregular lines, which can lead to gaps in performance compared to indirect representations such as segmentation-based or point-based methods. We have observed that these lanes are not intended to be irregular, but they appear zigzagged in the perspective view due to being drawn on uneven pavement. In this paper, we propose a new approach to the lane detection task by decomposing it into two parts: curve modeling and ground height regression. Specifically, we use a parameterized curve to represent lanes in the BEV space to reflect the original distribution of lanes. For the second part, since ground heights are determined by natural factors such as road conditions and are less holistic, we regress the ground heights of key points separately from the curve modeling. Additionally, we have unified the 2D and 3D lane detection tasks by designing a new framework and a series of losses to guide the optimization of models with or without 3D lane labels. Our experiments on 2D lane detection benchmarks (TuSimple and CULane), as well as the recently proposed 3D lane detection datasets (ONCE-3Dlane and OpenLane), have shown significant improvements. We will make our well-documented source code publicly available.

Motivation & Objective

• To address the performance gap in curve-based lane detection caused by irregular lane appearances in perspective views due to uneven pavement.
• To improve the holistic representation of lanes by modeling their intrinsic geometric shape in BEV space, separate from ground height variations.
• To unify 2D and 3D lane detection into a single framework using a shared 3D lane detection head and 2D supervision with regularization.
• To develop a decoupled head that separately regresses curve parameters and ground heights for better generalization and robustness.
• To achieve state-of-the-art performance on both 2D (TuSimple, CULane) and 3D (ONCE-3DLane, OpenLane) lane detection benchmarks.

Proposed method

• Decomposes lane detection into two stages: (1) curve modeling in BEV space using a parameterized polynomial (third-order) to represent the intrinsic lane shape.
• Separately regresses ground heights at key points using a dedicated head, treating height variations as local, non-holistic perturbations.
• Maps 3D lane coordinates (X, Z, height) back to 2D image space using camera intrinsic parameters to generate perspective-view predictions.
• Unifies 2D and 3D detection via a single 3D detection head, enabling end-to-end training with 2D supervision and a regularization loss to encourage flat ground height predictions.
• Uses a Curve/HeightFormer module to attend to spatial relationships in both curve and height representations, improving feature learning.
• Applies a regularization loss to constrain ground height predictions to be as flat as possible, allowing the model to focus on real road irregularities.

Experimental results

Research questions

• RQ1 Why do curve-based lane detection methods underperform on real-world data despite their holistic representation?
• RQ2 To what extent do ground surface irregularities, rather than lane geometry, cause fitting errors in perspective-view lane detection?
• RQ3 Can decoupling curve modeling from ground height regression improve lane detection accuracy and robustness?
• RQ4 How can 2D and 3D lane detection be unified under a single framework to leverage abundant 2D data for 3D performance?
• RQ5 What is the optimal curve representation and head design for achieving state-of-the-art performance in both 2D and 3D lane detection?

Key findings

• DecoupleLane achieves 75.07% F1 on the ONCE-3DLane validation set, outperforming all prior 3D lane detection methods.
• On OpenLane, DecoupleLane achieves 51.2% mF1, surpassing state-of-the-art methods like PersFormer (50.5%) and CurveFormer (50.5%).
• The ablation study shows that the DecoupleHead contributes the most to performance gain, with a 1.6% improvement over the baseline.
• Using a third-order polynomial for curve representation yields the best balance of performance and efficiency, outperforming second-order and fourth-order polynomials.
• Bézier curves do not improve performance over polynomials in the DecoupleHead, suggesting polynomials are more effective for this task.
• The unified training framework with 2D supervision and regularization enables strong 3D generalization, even without 3D lane labels during training.