DiffusionLane: Diffusion Model for Lane Detection

Introduction

• Learnable lane anchors can perform well on a specific dataset but generalize poorly under distribution shift, requiring retraining in those scenarios.DiffusionLane instead starts from random lane anchors and progressively refines them, avoiding learnable anchors and their training-dataset overfitting.Source: Introduction · p. 1
• The paper presents DiffusionLane as the first successful application of diffusion models to lane detection, extending a denoising approach previously explored in other perception tasks.The cited prior diffusion applications include object detection, tracking, and image segmentation.Source: Introduction · pp. 1–2
• DiffusionLane adds Gaussian noise to ground-truth lane starting points and angles during training, then denoises noisy anchors into target lanes; inference starts from Gaussian-sampled random anchors.Noisy anchors are used for RoI pooling to crop encoder features, which the decoder uses to predict lanes without noise.Source: Introduction · pp. 1–2, Figure 1 · p. 1
• The hybrid diffusion decoder combines global- and local-level decoding to refine higher-quality lane anchors, while a training-only auxiliary head uses learnable anchors for extra positive supervision.The auxiliary head is intended to improve encoder feature representation, which random anchors otherwise degrade.Source: Introduction · p. 2
• 81.32% F1 score on CULane marks a new state-of-the-art result; on MuLane, DiffusionLane reaches 86.23% accuracy, 1.21% above CLRerNet.The paper also reports state-of-the-art results on Carlane, Tusimple, CULane, and LLAMAS.Source: Introduction · p. 2

Lane Detection

• Prior methods divide into segmentation-based (SCNN, Pan et al. 2018; RESA, Zheng et al. 2021), anchor-based (UFLD, Qin, Wang, and Li 2020; Line-CNN, Li et al. 2019; CLRNet, Zheng et al. 2022), and parameter-based approaches (PolyLaneNet, Tabelini et al. 2021b; LSTR, Liu et al. 2021b).Segmentation methods treat lane detection as segmentation; anchor-based methods refine predefined anchors; parameter-based methods model lanes parametrically.Source: Lane Detection · p. 2

Diffusion Model

• Visual-perception diffusion has been used for segmentation (DDP, Ji et al. 2023), tracking (DiffusionTrack, Xie, Wang, and Ma 2024), box prediction (DiffusionDet, Chen et al. 2023a), and aerial lane-graph segmentation (LSR-DM, Ruiz et al. 2024).LSR-DM segments lane graphs from aerial imagery rather than detecting lanes from an on-board camera.Source: Diffusion Model · p. 2
• DiffusionLane is presented as the first use of diffusion denoising for traditional on-board lane detection.Prior visual diffusion methods add noise to bounding boxes or feature maps; DiffusionLane diffuses lane starting points and angles.Source: Diffusion Model · p. 2
• DiffusionLane also proposes hybrid decoding to reduce the influence of random lane anchors.This is identified as a difference from previous visual diffusion models.Source: Diffusion Model · p. 2

Preliminaries

• Each lane anchor samples evenly spaced points along the image height and is refined through foreground classification, parameter and length prediction, and point-offset regression.The model predicts background or foreground probabilities, then estimates the starting point, angle, length, and N horizontal offsets for foreground anchors.Source: Preliminaries · pp. 2–3

Diffusion Model for Lane Detection

• The image encoder and SAFPN produce multi-scale features once, while the hybrid diffusion decoder iterates T times on those features to refine noisy anchors, reducing repeated whole-model computation.A training-only auxiliary head uses learnable lane anchors to improve encoder feature representation.Source: Diffusion Model for Lane Detection · p. 3, Figure 3 · p. 4

Training

• Training pads each image to Ntrain lanes with random anchors, then corrupts ground-truth anchors by adding scheduled Gaussian noise to their starting-point coordinates and angle.The authors report that concatenating random anchors works best among the padding strategies considered.Source: Training · p. 3

Inference

• At inference, the decoder iteratively refines Gaussian-sampled anchors with DDIM, replacing filtered background anchors with random ones between steps before final NMS.Resampling is used to align the inference anchor distribution with the padded training setup.Source: Inference · p. 4

Hybrid Decoding

• To address weak encoder features from low-quality random anchors, the hybrid decoder combines global RoIGather with local self-attention and dynamic convolution across image-feature scales.The global decoder is primary; weighted additions inject local features into its inputs and outputs, and its output generates refined anchors.Source: Hybrid Decoding · pp. 4–5
• A training-only auxiliary head with learnable lane anchors enriches encoder supervision by providing targets for positive and negative predictions across multi-scale features.Its losses match those of the selected original head, such as CLRNet or CLRerNet.Source: Hybrid Decoding · p. 5

Datasets

• The experiments cover CULane, LLAMAS, Tusimple, and Carlane; Carlane contains TuLane, MoLane, and MuLane, each with over 20K training images.The datasets include varied scales and settings, including highway scenes on Tusimple and over 100K images on LLAMAS.Source: Datasets · p. 5

Evaluation Metrics

• CULane and LLAMAS are evaluated with F1, while Tusimple and Carlane use accuracy, false positives (FP), and false negatives (FN).Source: Evaluation Metrics · p. 5

Implementation Details

• The implementation uses T = 2, a CLRNet auxiliary head, a 0.4 foreground-background confidence threshold, initial noise scale 2, and three hybrid diffusion blocks.Source: Implementation Details · p. 5

Comparison with the State-of-the-art Methods

• DiffusionLane gains 2.74% (91.28% vs. 88.54%), 3.65% (91.00% vs. 87.35%), and 1.22% (86.23% vs. 85.01%) accuracy over CLRerNet on MoLane, TuLane, and MuLane.The paper attributes its domain-shift performance to Gaussian-sampled, domain-agnostic lane anchors.Source: Table 1 · p. 5, Comparison with the State-of-the-art… · p. 6, Table 2 · p. 6
• On LLAMAS, DiffusionLane raises F1 from 97.42% to 97.59% over Lane2Seq; on Tusimple, ResNet34 achieves 96.89% accuracy and ResNet101 lowers FP from 1.99% to 1.97%.The authors report that performance gaps among Tusimple methods are small because the dataset is small.Source: Comparison with the State-of-the-art… · p. 6, Table 3 · p. 6, Table 4 · p. 6
• On CULane, DiffusionLane with MobileNetV4-Hybrid-M reaches 81.32% F1, while ResNet34 improves over CLRerNet by 0.49% (80.24% vs. 79.73%).Qualitatively, DiffusionLane detects lanes under distribution shift and extreme lighting where CLRNet and Lane2Seq fail.Source: Comparison with the State-of-the-art… · p. 6, Figure 6 · p. 6

Ablation Study

• On CULane, diffusion raises F1 by 3.64% (78.38% vs. 74.74%), hybrid decoding adds 1.08% (79.46% vs. 78.38%), and auxiliary supervision raises it to 80.24% from 79.46%.Using random lane anchors degrades F1 from 79.96% to 74.74%; the auxiliary head supplies extra positive supervision, while hybrid decoding combines complementary decoders and refines anchor features.Source: Ablation Study · pp. 6–7, Table 6 · p. 7
• Ablations favor T = 2, DDIM with lane-anchor resampling, and Gaussian-distributed random-anchor padding; larger T degrades performance and adds computational burden.Source: Ablation Study · p. 7, Table 7 · p. 7, Table 8 · p. 7

Conclusion

• Experiments on four benchmarks show that DiffusionLane achieves favorable performance compared with existing lane detectors.Source: Conclusion · p. 7
• DiffusionLane casts lane detection as a diffusion process from noisy lane anchors to target lanes, using random lane anchors and denoising.Source: Conclusion · p. 7
• The random-anchor denoising paradigm enables DiffusionLane to handle distribution-shift scenarios without retraining.Source: Conclusion · p. 7
• Its hybrid decoding strategy combines a hybrid diffusion decoder and auxiliary detection head to improve lane-anchor refinement and feature representation.Source: Conclusion · p. 7

• It provides pseudocode for the training and inference stages.Source: Suppmentary Material for… · p. 10
• Additional experiments cover evaluation metrics, implementation details, and ablation studies.Source: Suppmentary Material for… · p. 10
• The material also describes the method’s limitations and broader impact.Source: Suppmentary Material for… · p. 10

Pseudo-codes of training and inference stage

• Algorithms 1 and 2 provide pseudocode for the training and inference stages, respectively.The text defines alpha cumprod(t) as the product of bar alpha_i for i=1 to t.Source: Pseudo-codes of training and inference… · p. 10
• In Algorithm 2, linespace means generating evenly spaced values.Source: Pseudo-codes of training and inference… · p. 10

Evaluation metrics

• CULane and LLAMAS use F1 score, while TuSimple and CArLane use accuracy, FP, and FN to evaluate performance.For the latter datasets, a lane point is counted correct when its distance is below t_pc = 20 cos(a_yl), where a_yl is the corresponding ground-truth angle.Source: Evaluation metrics · p. 10

Additional implementation details

• Experiments use ImageNet1K-pretrained ResNet, DLA, or MobileNetV4-Hybrid-M encoders and train with AdamW at learning rate 0.0003.Training uses a cosine learning-rate schedule, one 24-GB 3090 GPU, batch size 20, dataset-specific epoch counts, and N_train = 800.Source: Additional implementation details · p. 10

Additional ablation studies

• Noise scale 2 yields optimal performance in the tested ablation, exceeding the scales used for image generation (1) and segmentation (0.1).The cited prior-work markers are (Song, Meng, and Ermon 2020) and (Ji et al. 2023).Source: Additional ablation studies · p. 10, Table 9 · p. 10
• DiffusionLane prefers 800 lane anchors: fewer make target-near anchors sparse and optimization difficult, whereas larger counts increase computation.The study selects 800 anchors as a balance of performance and efficiency.Source: Additional ablation studies · p. 10, Table 10 · p. 10
• DiffusionLane lacks an inference-speed advantage, which the authors attribute possibly to multiple decoder runs.Inference speed was tested on a single 2080Ti GPU without TensorRT; improving speed is ongoing work.Source: Additional ablation studies · p. 10, Table 11 · p. 10
• Hybrid decoding consistently improves existing anchor-based lane detectors, supporting its use to optimize learnable lane-anchor parameters.The passage does not report numerical gains.Source: Additional ablation studies · p. 10, Table 12 · p. 11
• Using all lane points as the diffusion target requires more training epochs and computation than diffusing the starting point and angle, so the latter are selected.The comparison indicates that optimizing all lane points is harder.Source: Additional ablation studies · pp. 10–11, Table 13 · p. 11

Limitation

• Inference speed is a major limitation because DiffusionLane uses dense lane anchors and runs the decoder multiple times.The authors identify efficient decoding as future work.Source: Limitation · p. 11

Broader Impact

• The authors propose diffusion with random lane anchors as a possible lane-detection model that generalizes to distribution-shift scenarios without retraining.They hope DiffusionLane can serve as a simple, effective baseline for anchor-based methods and inspire higher-performance, more efficient architectures.Source: Broader Impact · p. 11