[Paper Review] BSNet: Lane Detection via Draw B-spline Curves Nearby

Haoxin Chen, Mengmeng Wang+1 more|arXiv (Cornell University)|Jan 17, 2023

Autonomous Vehicle Technology and Safety10 citations

TL;DR

BSNet uses clamped quasi-uniform b-spline curves to represent lanes, with a light proposal-based network and a novel curve distance, achieving state-of-the-art results on Tusimple, CULane, and LLAMAS while running at 197 FPS.

ABSTRACT

Curve-based methods are one of the classic lane detection methods. They learn the holistic representation of lane lines, which is intuitive and concise. However, their performance lags behind the recent state-of-the-art methods due to the limitation of their lane representation and optimization. In this paper, we revisit the curve-based lane detection methods from the perspectives of the lane representations' globality and locality. The globality of lane representation is the ability to complete invisible parts of lanes with visible parts. The locality of lane representation is the ability to modify lanes locally which can simplify parameter optimization. Specifically, we first propose to exploit the b-spline curve to fit lane lines since it meets the locality and globality. Second, we design a simple yet efficient network BSNet to ensure the acquisition of global and local features. Third, we propose a new curve distance to make the lane detection optimization objective more reasonable and alleviate ill-conditioned problems. The proposed methods achieve state-of-the-art performance on the Tusimple, CULane, and LLAMAS datasets, which dramatically improved the accuracy of curve-based methods in the lane detection task while running far beyond real-time (197FPS).

Motivation & Objective

• Motivate curve-based lane detection by improving globality and locality in lane representations.
• Propose a b-spline curve representation that supports flexible fitting and efficient optimization.
• Design BSNet to aggregate global and local features for accurate lane proposals.
• Introduce a robust curve distance that mitigates ill-conditioned optimization in curve fitting.

Proposed method

• Adopt clamped quasi-uniform b-spline curves as the lane representation to balance locality and globality.
• Develop a proposals features init module to generate lane proposals from a ResNet+FPN backbone.
• Incorporate local features by sampling points on proposals and applying self-attention to produce F_all.
• Define a new curve distance that computes D_AB as D_B→A + D_A→B using point-to-curve distances with normalization.
• Use a simple prediction head with Fast NMS and a loss that includes regression, length, start-point, and focal classification terms.
• Train with data augmentation on three datasets (CULane, Tusimple, LLAMAS) using AdamW and cosine learning rate scheduling.

Experimental results

Research questions

• RQ1 Can a b-spline-based lane representation achieve both globality and locality to improve lane fitting, especially for curvy lanes?
• RQ2 Does a dedicated curve distance address ill-conditioned optimization issues in curve-based lane detection?
• RQ3 Can a lightweight BSNet with global/local features reach state-of-the-art accuracy while maintaining high inference speed across diverse datasets?
• RQ4 How does the number of control points and sample points affect fitting quality and stability?

Key findings

• BSNet achieves state-of-the-art results on CULane, Tusimple, and LLAMAS datasets among curve-based methods.
• ResNet-18/34 BSNet variants reach up to 197 FPS on GTX 1080Ti, with competitive or superior accuracy compared to stronger backbones.
• B-spline representation yields better fitting than Bézier in ablations, especially when aided by local features.
• The proposed curve distance alleviates ill-conditioned optimization and improves regression stability and convergence.
• A simple network design with a small detection head provides strong performance and ease of multitasking integration.