Source-linked AI summary
CurveLane-NAS: Unifying Lane-Sensitive Architecture Search and Adaptive Point Blending
Hang Xu, Shaoju Wang, Xinyue Cai, Wei Zhang, Xiaodan Liang, Zhenguo Li
TL;DR
Curve lane detection challenges hand-designed methods because remote curves require both long-range context and detailed trajectories. CurveLane-NAS jointly searches lane-sensitive architectures and adaptive point blending, remaining resilient on CurveLanes while achieving 74.8% F1-score on CULane.
Problem
Curve lane detection is difficult for hand-designed methods because remote curve parts require both long-range contextual information and detailed curve trajectories.
Method
CurveLane-NAS unifies feature fusion search, elastic backbone search, and adaptive point blending to combine multi-level predictions for curve-sensitive detection.
Results
The method remains resilient on CurveLanes, where state-of-the-art methods suffer substantial performance drops, and achieves a new 74.8% F1-score on CULane.
Takeaways & Limitations
CurveLane-NAS provides an efficient speed/FLOPS trade-off, while CurveLanes offers a challenging benchmark with more curve lanes per image.
Takeaways & Limitations
Existing Line-NMS cannot handle predictions from multiple feature levels equally, and the masking design relies on a conjecture about low-level features.
Abstract
from arXiv · showhide
We address the curve lane detection problem which poses more realistic challenges than conventional lane detection for better facilitating modern assisted/autonomous driving systems. Current hand-designed lane detection methods are not robust enough to capture the curve lanes especially the remote parts due to the lack of modeling both long-range contextual information and detailed curve trajectory. In this paper, we propose a novel lane-sensitive architecture search framework named CurveLane-NAS to automatically capture both long-ranged coherent and accurate short-range curve information while unifying both architecture search and post-processing on curve lane predictions via point blending. It consists of three search modules: a) a feature fusion search module to find a better fusion of the local and global context for multi-level hierarchy features; b) an elastic backbone search module to explore an efficient feature extractor with good semantics and latency; c) an adaptive point blending module to search a multi-level post-processing refinement strategy to combine multi-scale head prediction. The unified framework ensures lane-sensitive predictions by the mutual guidance between NAS and adaptive point blending. Furthermore, we also steer forward to release a more challenging benchmark named CurveLanes for addressing the most difficult curve lanes. It consists of 150K images with 680K labels.The new dataset can be downloaded at github.com/xbjxh/CurveLanes (already anonymized for this submission). Experiments on the new CurveLanes show that the SOTA lane detection methods suffer substantial performance drop while our model can still reach an 80+% F1-score. Extensive experiments on traditional lane benchmarks such as CULane also demonstrate the superiority of our CurveLane-NAS, e.g. achieving a new SOTA 74.8% F1-score on CULane.
1 Introduction
Curve lane detection is difficult because remote curve sections require both long-range context and detailed trajectory modeling, while existing datasets underrepresent such lanes. CurveLane-NAS addresses these limitations with unified architecture and post-processing search, and CurveLanes provides a challenging benchmark for evaluation.
- Curve lanes are difficult because their varied shapes, occlusions, increased remote-part curvature, and harsh conditions challenge detection.
- 2.1% of CULane images and 30% of TuSimple images contain curve lanes, limiting these datasets for measuring curve lane detection.
- CurveLanes contains 150K images and 680K annotated curve-lane labels, with more than 135K images containing at least one curve lane.
- Existing dense methods struggle with remote curve parts and fixed lane counts, while proposal-based methods lack flexible curvature modeling or multi-scale context.
- CurveLane-NAS searches elastic backbones, multi-level feature fusion, and adaptive point blending to combine long-range coherent and short-range curve information.
- 74.8% F1-score on CULane is reported for the searched model, with reduced FLOPS relative to competing methods.
2 Related Work
Lane detection research has progressed from handcrafted features to deep dense-prediction and proposal-based methods, while NAS automates architecture design through evolutionary or gradient-based search.
- Lane detection localizes the accurate location and shape of road lanes and is a core problem in assisted and autonomous driving systems.
- Figure 2 contrasts dense prediction, proposal-based detection, and CurveLane-NAS as a unified search framework for holistic lane architectures.
- Dense prediction methods formulate lane detection as semantic segmentation, assigning each image pixel a lane or non-lane label.
- Neural Architecture Search automatically finds task- and dataset-specific network architectures instead of relying entirely on expert design.
- NAS search strategies include evolutionary architecture mutation and gradient-based continuous relaxation with weight-sharing and differentiable optimization.
3 CurveLane-NAS framework
CurveLane-NAS unifies architecture search and post-processing to capture coherent long-range lane context and detailed short-range curve information. Its search modules jointly optimize feature extraction, feature fusion, and point-level refinement for accuracy and efficiency.
- CurveLane-NAS combines elastic backbone, feature fusion, and adaptive point blending search in one lane-sensitive framework.The unified search also optimizes post-processing for multi-level lane predictions.
- 3.1 Elastic Backbone Search Module: The elastic backbone search explores task-specific network width, depth, channel increases, down-sampling locations, and computational allocation.Its search space includes 3 or 4 stages, base channel sizes from 48 to 128, 10 to 45 blocks, and about 5 × 10^12 choices.
- 3.4 Unified Multi-objective Search: A unified multi-objective search algorithm generates a Pareto front trading prediction accuracy against FLOPS.The searched design aims to allocate computation across receptive fields and spatial resolutions for efficient lane detection.
- 3.2 Feature Fusion Search Module: Feature fusion search combines multi-level features and selects prediction-head levels to balance long-range coherence with short-range curve detail.Fusion layers resample and concatenate two feature levels, while each feature level may receive a prediction head.
- 3.3 Adaptive Point Blending Search Module: Point blending replaces selected points in high-confidence lane predictions with accurate local points to refine remote and detailed curve shapes.The procedure follows score modification, low-score filtering, and NMS grouping before blending prediction points.
- 3.3 Adaptive Point Blending Search Module: Adaptive score masking assigns different regions of interest across feature levels because low-level and high-level predictions differ in local accuracy and long-range consistency.The masking formulation is motivated by the conjecture that low-level features may perform better in remote lane regions near the image center.
4 Experiments
Experiments evaluate CurveLane-NAS on CurveLanes, CULane, and TuSimple, using dataset-specific official metrics and searched models with different computational constraints. The method achieves strong curve-lane performance, CULane state-of-the-art results, and transferability across lane-detection tasks.
- 4.1 New CurveLanes Benchmark: CurveLanes contains 150K images with 650K annotated lane labels, including more lanes, higher resolution, and more curved scenes than existing benchmarks.The dataset targets difficult scenarios such as curves and multi-lane traffic scenes; its split is 100K training, 20K validation, and 30K testing images.
- 4.2 Other Datasets and Evaluation Metrics: Evaluation uses IoU-based F1-measure for CULane and CurveLanes, while TuSimple uses its official lane-point accuracy metric.For CULane and CurveLanes, predictions with IoU above 0.5 count as true positives; TuSimple accuracy is based on correctly predicted lane points.
- 4.3 Lane Detection Results: CurveLane-L achieves 74.8 F1 on CULane, while CurveLane-M improves F1 by 1.9 over SCNN with 9x fewer FLOPS and by 1.5 over R101-SAD with 4.5x fewer FLOPS.The searched models are denoted CurveLane-S, CurveLane-M, and CurveLane-L under different computational constraints.
- 4.3 Lane Detection Results: On CurveLanes, state-of-the-art methods lose 20%–30% F1, whereas CurveLane-S reaches 81.12% F1 and exceeds SCNN by 16%.Qualitative results show stronger performance in large curves, night scenes, and wet roads.
- 4.3 Lane Detection Results: Architectures searched on CULane transfer to TuSimple with performance comparable to SCNN but substantially faster.The study searches on CULane and CurveLanes, then evaluates transferability on TuSimple.
- 4.4 Ablative Study: Adaptive Point Blending improves curve and remote-lane predictions over Plain-NMS, while multi-level heads help smaller models more and blending benefits larger models more.The ablation study is conducted on CULane using searched and ResNet101 backbones.
5 Conclusion
CurveLane-NAS unifies lane-sensitive architecture search with adaptive point blending for curve lane detection. It captures long-range and short-range curve information, allocates computation efficiently, and introduces CurveLanes as a challenging benchmark.
- 5 Conclusion: CurveLane-NAS unifies lane-sensitive architecture search and adaptive point blending for curve lane detection.The framework automatically fuses and captures long-ranged coherent information and accurate curve information.
- 5 Conclusion: The searched networks achieve state-of-the-art speed/FLOPS trade-offs compared with existing methods.
- 5 Conclusion: CurveLanes provides a larger, more curve-rich benchmark for difficult lane-detection scenarios.The dataset contains more images and more lanes per image than existing benchmarks.