Skip to content

title: >- [Paper Note] SGC-Lane: Monocular 3D Lane Detection with Standard-Definition Map Guidance and Lane Completion description: >- [ECCV 2026][autonomous_driving][3D Lane Detection] SGC-Lane combines SD map guidance and step-consistency lane completion to tackle ill-posed monocular 3D lane detection. tags: - ECCV 2026 - autonomous_driving - 3D Lane Detection - SD Map Guidance - Lane Completion date: 2026-09-19 content_hash: 012e3987b8f6abaa

SGC-Lane: Monocular 3D Lane Detection with Standard-Definition Map Guidance and Lane Completion

Conference: ECCV 2026
Paper: ECCV Original
Code: https://github.com/FuqingJIang/SGC-Lane
Area: Autonomous Driving
Keywords: 3D Lane Detection, Monocular Vision, Map Prior, Lane Completion, Sequence Consistency

TL;DR

To overcome the ill-posed nature of monocular 3D geometry reasoning and endpoint truncation in sparse-point representations, SGC-Lane adapts coarse Standard-Definition (SD) road-level centerlines into lane-level probability priors via front-view width prediction and introduces a step-consistency completion head (SCC-Head) with endpoint clustering, achieving a state-of-the-art 65.6% F1-score on OpenLane.

Background & Motivation

Monocular 3D lane detection provides an economically attractive perception foundation for autonomous vehicles, yet reconstructing accurate 3D physical geometry from a single 2D image is inherently an ill-posed inverse problem. Visual evidence is frequently ambiguous or incomplete in everyday driving scenarios featuring long-range horizons, sharp curves, steep grade variations, and severe vehicle occlusions. While high-definition (HD) maps supply abundant geometric and topological priors beyond the visual line of sight, their formidable construction, licensing, and real-time maintenance costs severely hinder widespread deployment. Conversely, standard-definition (SD) navigation maps offer ubiquitous global coverage and low cost, but their direct exploitation in 3D physical lane detection remains fundamentally hindered: SD maps only document road-level centerlines—abstracting an entire multi-lane corridor into a solitary curve with lateral inaccuracies of several meters—resulting in a severe representation gap against concrete lane boundaries.

Simultaneously, the mainstream sparse-point paradigm pre-samples points along the longitudinal depth axis to represent elongated lanes. Model predictions frequently suffer from artificial truncation and broken fragments at endpoints due to horizon limits and occlusion during training annotation. Recent patching remedies like EP-Head regress distances to endpoints independently per anchor, leaving predictions vulnerable to localized point errors and overlooking the structural regularity that sparse-point representations inherently possess along the longitudinal dimension.

To reconcile coarse topological map priors with fine-grained physical lane boundaries while resolving fragmented sparse-point detections, this paper integrates map guidance and sequence-consistent lane completion. Core idea: dynamically predict road widths from front-view features to expand coarse SD road centerlines into parallel lane-boundary probability maps that guide visual attention and local spatial cross-attention query initialization, combined with a step-consistency completion head that enforces second-order difference constraints along the longitudinal axis for robust endpoint clustering.

Method

Overall Architecture

SGC-Lane consists of two coordinated stages. In the first stage, the FV-Guided Map Encoder aligns road-level SD centerlines with front-view image context, expanding them into candidate lane boundaries to generate a lane probability map and lane-aware map features. In the second stage, the Local Spatial Cross-Attention (LSCA) module injects localized map priors into lane queries, followed by a LATR-style transformer decoder equipped with a perspective-aware non-uniform sampling plane (NUS-Plane) and a Step Consistency Completion Head (SCC-Head) that outputs fully completed 3D lanes.

The input comprises a monocular front-view image and a set of SD centerlines \(\mathcal{M} = \{C_i\}_{i=1}^N\) clipped within a spatial ground window (\(x \in [-15\,\text{m}, 15\,\text{m}], y \in [3\,\text{m}, 103\,\text{m}]\)). The FV-Guided Map Encoder predicts a dynamic width \(\hat{w}_i\) for each centerline using pooled front-view features, offsets the centerline along its normal vectors to produce \(\beta\) parallel lane boundaries, and projects them onto the image plane to construct a rasterized lane probability map \(\mathcal{H}\). This probability map modulates front-view features through a learnable gate. Subsequently, per-point lane queries estimate their coarse ground coordinates and retrieve the \(K\) nearest prior map tokens via LSCA. The refined queries are decoded layer-by-layer, and the SCC-Head predicts endpoint offsets under second-order difference consistency, which are then clustered to reconstruct continuous 3D lane geometries.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    InImg["Front-View Image"]
    InMap["SD Map Road Centerlines"]

    subgraph S1["FV-Guided Map Encoder"]
        direction TB
        FVE["FV Feature & Context Extraction"]
        CW["Dynamic Lane Width Prediction<br/>Normal Expansion into β Lane Boundaries"]
        Heatmap["Ground-to-Image Projection & Rasterization<br/>Generate Lane Probability Map H"]
        MapFeat["Coordinate & Positional Encoding<br/>Lane-Aware Map Features M_map"]
        FVE --> CW
        CW --> Heatmap
        CW --> MapFeat
    end

    InImg --> FVE
    InMap --> CW

    Gate["Learnable Gated Weighting<br/>Modulate Front-View Features"]
    Heatmap --> Gate
    InImg --> Gate

    subgraph S2["Local Spatial Cross-Attention (LSCA)"]
        direction TB
        QGen["Per-Point Query Assembly<br/>Q_lane ⊕ Q_point"]
        Loc["Ground Coordinate Prediction (x, y)"]
        TopK["Top-K Local Spatial Token Retrieval"]
        CA["Cross-Attention Prior Fusion"]
        QGen --> Loc --> TopK --> CA
    end

    Gate --> QGen
    MapFeat --> TopK
    MapFeat --> CA

    Dec["LATR Decoder Refinement<br/>with Perspective-Aware NUS-Plane"]
    CA --> Dec

    subgraph S3["Step Consistency Completion Head (SCC-Head)"]
        direction TB
        OffPred["Per-Anchor Endpoint Offsets (s_k, e_k)"]
        ConsLoss["Second-Order Step Difference Consistency"]
        Clust["Endpoint Candidate Clustering"]
        OffPred --> ConsLoss
        OffPred --> Clust
    end

    Dec --> OffPred
    Clust --> Out["Final Complete 3D Lanes"]

Key Designs

1. FV-Guided Map Encoder: Bridging Road-Level Topologies to Lane-Level Priors SD map centerlines capture macro road topology but lack lane boundary delineations and exhibit several meters of lateral offset error. To bridge this representation divide, the module encodes sampled points \(p_{i,j}\) of the \(i\)-th centerline into point-wise embeddings \(f_{i,j}\), aggregating them via average pooling into a centerline embedding \(\bar{F}_i\). Simultaneously, spatial average pooling extracts a global context vector \(v_{\text{ctx}}\) from front-view features \(F_{\text{fv}}\). A lightweight MLP then predicts the lane width: $\(\hat{w}_i = \text{MLP}(\phi(\bar{F}_i, v_{\text{ctx}}))\)$ Using \(\hat{w}_i\), the centerline is translated along its unit normal vectors to generate \(\beta\) parallel lane-level centerlines, which are further offset by \(\pm \frac{1}{2}\hat{w}_i\) to construct physical left and right boundary sets. Projecting these boundaries onto the front-view plane and applying Gaussian rasterization produces a normalized lane probability map \(\mathcal{H} \in [0, 1]^{H \times W \times 1}\). This map applies a residual gate \(\tilde{\mathbf{F}}_{\text{fv}} = \mathbf{F}_{\text{fv}} + \alpha \cdot \text{proj}(\mathcal{H})\) to focus image features on plausible lane regions, while coordinate and sequence positional embeddings form structured lane-aware map features \(\mathbf{M}_{\text{map}}\).

2. Local Spatial Cross-Attention: Spatially Constrained Prior Injection Direct global cross-attention between queries and all map tokens causes long-range feature confusion and high computational overhead. LSCA adopts a localized retrieval strategy. The query generator combines instance queries \(\mathbf{Q}_{\text{lane}}\) and point embeddings \(\mathbf{Q}_{\text{point}}\) into per-point queries \(q_p \in \mathbb{R}^C\). For each query, a compact MLP predicts its ground lateral position \(\hat{x}_p\), paired with its preset longitudinal anchor coordinate \(\hat{y}_p = y_{\text{anchor}}[p]\). The module then queries the ground coordinates \(\mathbf{C}_{\text{map}}\) of map features to isolate the \(K\) nearest tokens: $\(\text{idx}_p = \operatorname{TopK}_K \big(-\|\mathbf{C}_{\text{map}} - (\hat{x}_p, \hat{y}_p)\|_2\big)\)$ Cross-attention is executed strictly over these \(K\) retrieved tokens, injecting precise geometric guidance while shielding queries from irrelevant distant map segments.

3. Step Consistency Completion Head: Sequence-Constrained Endpoint Clustering In sparse-point lane representations, anchors are sampled at fixed intervals along the longitudinal (\(Y\)) axis. Consequently, the change in distance from consecutive anchors to the true endpoints should remain constant. Unlike prior works that regress endpoint distances independently per anchor, SCC-Head enforces a structural second-order difference consistency constraint on the predicted endpoint offsets \(\hat{s}_k\) and \(\hat{e}_k\): $\(\mathcal{L}_{\text{sy-cons}} = \frac{1}{M-2} \sum_{k=2}^{M-1} \left| \Delta \hat{s}^y_k - \Delta \hat{s}^y_{k-1} \right|, \quad \Delta \hat{s}^y_k = \hat{s}^y_{k+1} - \hat{s}^y_k\)$ At inference, each anchor projects its candidate endpoints \(p_k + \hat{s}_k\) and \(p_k + \hat{e}_k\). A lightweight clustering operator \(\mathcal{K}\) aggregates these candidates, taking the cluster centers as the definitive endpoints to synthesize missing intermediate points along the preset \(Y\)-sampling grid, eliminating horizon truncation.

Loss & Training

The overall framework is trained end-to-end with multi-task objectives encompassing lane classification, 3D point coordinate regression, and the SCC-Head objective: $\(\mathcal{L}_{\text{SCC}} = \mathcal{L}_{\text{base}} + \lambda_{\text{sy}}\mathcal{L}_{\text{sy-cons}} + \lambda_{\text{ey}}\mathcal{L}_{\text{ey-cons}}\)$ where \(\mathcal{L}_{\text{base}}\) denotes the L1 distance between predicted and ground-truth endpoint offsets, with hyper-parameters \(\lambda_{\text{sy}} = 1.0\) and \(\lambda_{\text{ey}} = 1.0\). The model employs a ResNet-50 backbone with input resolution resized to \(720 \times 960\). Training runs for 24 epochs using AdamW with weight decay 0.01 and an initial learning rate of \(2 \times 10^{-4}\) under a cosine annealing schedule across 8 NVIDIA RTX 4090 GPUs with a global batch size of 32.

Key Experimental Results

Main Results

On the large-scale OpenLane validation benchmark, SGC-Lane outperforms existing state-of-the-art monocular 3D lane detection methods (corresponding to Table 1 in the paper):

Methods Backbone F1 (%) ↑ Near X Error (m) ↓ Far X Error (m) ↓ Near Z Error (m) ↓ Far Z Error (m) ↓
PersFormer (ECCV 2022) EfficientNet 50.5 0.485 0.553 0.364 0.413
Anchor3DLane (CVPR 2023) ResNet-18 53.1 0.300 0.311 0.103 0.139
BEV-LaneDet (CVPR 2023) ResNet-34 58.4 0.309 0.659 0.244 0.631
LATR (ICCV 2023) ResNet-50 61.9 0.219 0.259 0.075 0.104
LaneCPP (CVPR 2024) EfficientNet 60.3 0.264 0.310 0.077 0.177
GLane3D (CVPR 2025) ResNet-50 63.9 0.193 0.234 0.065 0.090
SC-Lane (ICCV 2025) ResNet-50 64.3 0.227 0.251 0.088 0.128
Chang et al. / EP-Head (CVPR 2025) ResNet-50 64.7 0.205 0.255 0.074 0.105
SGC-Lane (Ours) ResNet-50 65.6 0.204 0.254 0.076 0.107
GLane3D (CVPR 2025) Swin-B 66.0 0.170 0.203 0.063 0.087
SGC-Lane (Ours) Swin-B 68.0 0.197 0.237 0.072 0.101

Scenario-wise evaluations (Table 2 in the paper) indicate clear advantages in topologically demanding scenarios: Up&Down reaches 56.9% F1 (+1.6%), Curve achieves 72.4% (+0.9%), Intersection attains 57.1% (+1.0%), Merge&Split hits 65.5%, and Night conditions register 58.6%.

Ablation Study

The modular ablation on OpenLane confirms the progressive contributions of SD map guidance, SCC-Head, and the non-uniform sampling plane (corresponding to Table 3 in the paper):

Exp SD-Map SCC-Head NUS-Plane F1 (%) ↑ Gain (%) ↑ Near X Error (m) ↓ Far X Error (m) ↓ Near Z Error (m) ↓ Far Z Error (m) ↓ Note
1 - - - 61.9 - 0.219 0.259 0.075 0.104 Baseline LATR
2 ✓ - - 63.4 +1.5 0.208 0.241 0.072 0.103 SD Map prior only
3 - ✓ - 64.4 +2.5 0.231 0.284 0.079 0.111 SCC-Head only
4 ✓ ✓ - 65.3 +3.4 0.217 0.256 0.078 0.108 Map + Completion
5 - ✓ ✓ 64.7 +2.8 0.233 0.275 0.078 0.109 Completion + NUS-Plane
6 ✓ ✓ ✓ 65.6 +3.7 0.204 0.254 0.076 0.107 Full model

Furthermore, a direct head-to-head comparison on the LATR baseline (Table 4) shows that LATR alone achieves 61.9% F1; adding EP-Head yields 63.1% F1; while replacing it with SCC-Head raises the F1 to 64.4% (+1.3% over EP-Head), verifying the efficacy of second-order sequence consistency.

Key Findings

  • Introducing the SD map prior alone (Exp 2) increases the F1-score by 1.5% while reducing both lateral and vertical errors (near X error drops from 0.219m to 0.208m), proving that soft gating and local attention filter out map inaccuracies while preserving topology.
  • Applying SCC-Head alone (Exp 3) brings a substantial 2.5% F1 gain by recovering truncated segments, but slightly increases localization errors due to challenging extrapolated endpoints. Combining SD map guidance with SCC-Head (Exp 4) constrains the completion search space, achieving 65.3% F1 while returning localization error back to baseline levels (0.217m), highlighting strong synergy.
  • Performance gains are largest in complex geometry scenarios (slopes, sharp turns, intersections), validating that map priors effectively overcome the sight-distance limits of monocular cameras.

Highlights & Insights

  • Adaptive expansion from road-level curves to lane boundaries: Leveraging global front-view features to dynamically predict lane width \(\hat{w}_i\) resolves the fundamental representation discrepancy between SD navigation maps and physical lane markings.
  • Structural sequence constraint via second-order step differences: Exploiting the fixed longitudinal sampling interval of sparse-point representations through second-order differences regularizes endpoint regression without prescribing rigid step sizes.
  • Lightweight synergy of non-uniform planes and local attention: Combining a perspective-aware ground sampling plane with Top-K local spatial cross-attention injects dense 3D cues with minimal computational overhead.

Limitations & Future Work

  • Reliance on SD map availability: The framework assumes the presence of rough SD road-level centerlines. In completely unmapped rural or private areas, guidance gains may degrade.
  • Fragmented line detection in extreme weather: In heavy fog or blinding blizzards, F1 slightly trails bottom-up keypoint clustering methods like GLane3D, which better assemble severely broken markings.
  • Future directions: The authors plan to integrate temporal multi-frame feature fusion and explore explicit monocular depth priors to further enhance 3D stability.
  • vs LATR (ICCV 2023): LATR serves as the underlying transformer baseline without map priors or completion heads. SGC-Lane builds upon it, boosting F1 by 3.7% under the same ResNet-50 backbone.
  • vs EP-Head / Chang et al. (CVPR 2025): EP-Head pioneered endpoint regression for sparse points but treats anchors independently; SCC-Head introduces second-order sequence consistency and clustering, achieving a 1.3% higher F1 gain under identical baseline configurations.
  • vs SMERF / SEPT (2024-2025): While prior SD-map methods primarily target BEV centerline generation, SGC-Lane is the first to achieve end-to-end adaptive mapping from road centerlines to physical 3D lane boundaries in monocular perception.

Rating

  • Novelty: ⭐⭐⭐⭐☆ (Creative dynamic width expansion of SD maps and structurally sound sequence-consistent endpoint completion)
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Comprehensive evaluation on OpenLane, complete sub-scenario breakdown, systematic ablations, and direct head comparisons)
  • Writing Quality: ⭐⭐⭐⭐⭐ (Clear problem definition, mathematically coherent formulations, and well-structured diagrams)
  • Value: ⭐⭐⭐⭐☆ (Offers an accessible, practical avenue for commercial autonomous driving systems to exploit inexpensive navigation maps)