[Paper Review] ADNet: Lane Shape Prediction via Anchor Decomposition

Lingyu Xiao, Xiang Li+2 more|arXiv (Cornell University)|Aug 21, 2023

Autonomous Vehicle Technology and Safety4 citations

TL;DR

ADNet proposes a novel anchor decomposition approach for lane shape prediction by decoupling anchor learning into start point heatmaps and directional predictions, enhancing flexibility and quality. It integrates Large Kernel Attention in FPN for better context capture and introduces General Line IoU (GLIoU) loss, achieving state-of-the-art performance on VIL-100 and competitive results on CULane and TuSimple.

ABSTRACT

In this paper, we revisit the limitations of anchor-based lane detection methods, which have predominantly focused on fixed anchors that stem from the edges of the image, disregarding their versatility and quality. To overcome the inflexibility of anchors, we decompose them into learning the heat map of starting points and their associated directions. This decomposition removes the limitations on the starting point of anchors, making our algorithm adaptable to different lane types in various datasets. To enhance the quality of anchors, we introduce the Large Kernel Attention (LKA) for Feature Pyramid Network (FPN). This significantly increases the receptive field, which is crucial in capturing the sufficient context as lane lines typically run throughout the entire image. We have named our proposed system the Anchor Decomposition Network (ADNet). Additionally, we propose the General Lane IoU (GLIoU) loss, which significantly improves the performance of ADNet in complex scenarios. Experimental results on three widely used lane detection benchmarks, VIL-100, CULane, and TuSimple, demonstrate that our approach outperforms the state-of-the-art methods on VIL-100 and exhibits competitive accuracy on CULane and TuSimple. Code and models will be released on https://github.com/ Sephirex-X/ADNet.

Motivation & Objective

• Address the rigidity of fixed anchor-based lane detection methods that assume lane lines originate only from image edges.
• Improve anchor quality by replacing dispersed anchor prediction with a concentrated, learnable start point and direction prediction mechanism.
• Enhance feature representation for slender, continuous lane lines by extending receptive field using Large Kernel Attention (LKA) in Feature Pyramid Networks.
• Overcome limitations of standard IoU losses in complex scenarios by proposing a new General Line IoU (GLIoU) loss tailored for anchor-based lane detection.
• Develop a unified framework that maintains high accuracy and efficiency across diverse datasets, including VIL-100, CULane, and TuSimple.

Proposed method

• Decompose anchors into two components: a learnable heat map for start point locations and a directional map for anchor orientation, enabling global, flexible anchor generation.
• Introduce the Start Point Generate Unit (SPGU) to predict start points via a probability heatmap, decoupling anchor location from fixed image edges.
• Integrate Large Kernel Attention (LKA) into the Feature Pyramid Network (FPN) to expand the receptive field and improve context modeling for long, continuous lane lines.
• Propose the Adaptive Lane Aware Unit (ALAU) that uses a guidance map from start points and directions to dynamically adjust kernel offsets in feature aggregation.
• Design the General Line IoU (GLIoU) loss to better handle complex lane configurations, especially when anchors are not aligned with standard edge assumptions.
• Combine SPGU, ALAU, LKA, and GLIoU in an end-to-end training pipeline, with ablation studies validating each component’s contribution.

Experimental results

Research questions

• RQ1 Can anchor-based lane detection be made more flexible by decoupling start point prediction from fixed image edges?
• RQ2 Does a concentrated, learnable anchor proposal mechanism improve both anchor quality and detection accuracy compared to dispersed anchor prediction?
• RQ3 Can large receptive fields, enabled by Large Kernel Attention (LKA), significantly enhance feature representation for long, continuous lane lines?
• RQ4 Does a task-specific loss like General Line IoU (GLIoU) outperform standard IoU-based losses in complex, non-standard lane configurations?
• RQ5 To what extent do the proposed components—SPGU, ALAU, LKA, and GLIoU—collectively improve performance across diverse benchmarks like VIL-100, CULane, and TuSimple?

Key findings

• ADNet achieves a 77.56% F1@50 score on CULane, a 5.39% improvement over the baseline, with the addition of all components including SPGU, ALAU, GLIoU, and LKA.
• On the VIL-100 benchmark, ADNet achieves 90.39% F1@50 with ResNet-34 and LKA, a 1.17% improvement over the baseline, outperforming all existing state-of-the-art methods.
• The SPGU component alone contributes a 4.30% increase in F1@50 on CULane, demonstrating its critical role in enabling flexible anchor generation.
• The GLIoU loss provides a 5.21% performance boost on VIL-100 when using ResNet-18, indicating its effectiveness in complex scenarios where standard IoU losses fail.
• LKA significantly improves performance across both CULane and VIL-100, outperforming the baseline attention module and showing consistent gains on different backbones.
• Ablation studies confirm that SPGU has the largest impact, followed by ALAU and LKA, while GLIoU has a smaller but measurable effect, especially on VIL-100.