PGU-Decoder: Prior-Guided Unified Decoding for Industrial Road-Network Map Updating

Yuguang Ning, Bingbing Kang, Nan Zhao, Zhixuan Li, Ze Zhao, Dongdong Liu, Wenjian Zhu, Zekai Liu, Junhui WangarXiv 2026 · arXiv

TL;DR

PGU-Decoder is a unified, production-grade framework that automates the three most labor-intensive tasks in road-network map updating: topology generation, geometric correction, and attribute assignment. By leveraging existing map data as "geometric priors" and fixing reference points during decoding, it achieves over 98% accuracy, slashes manual verification costs by 95%, and enables map updates to move from a weekly to an hourly cadence at industrial scale.

What Problem Does This Paper Address?

Maintaining accurate digital maps for a billion users is a massive undertaking. Industrial map data is complex, spanning road-level connectivity (topology), precise lane configurations (geometry), and driving constraints like "drivable areas" (attributes).

Traditionally, these three dimensions are handled by separate, fragmented pipelines. This leads to:

  1. High Operational Costs: Each pipeline requires its own rule maintenance, quality assurance, and manual verification.
  2. Inefficiency: Manual construction of complex intersections can cost ~$3.50 (CNY 25) each, and verifying lane attributes costs ~$1.70 (CNY 12) per kilometer.
  3. Generalization Issues: Rule-based systems struggle with the diverse road configurations found across hundreds of different cities.

The authors aim to replace these siloed systems with a single, unified deep learning architecture that can handle all three tasks efficiently.

Methods

The core innovation of PGU-Decoder is its Prior-Guided Unified Decoding architecture, which is built upon Deformable DETR but modified for the specific constraints of map updating.

  • Prior Query Module (PQM): Instead of starting with random "learnable" queries, PQM samples points directly from existing (upstream) road geometries. These points are encoded into "prior queries" that tell the model exactly where to look.
  • Fixed Reference Points: In standard object detection models, the decoder iteratively refines where it is looking. PGU-Decoder argues that since we already have coarse map data, we should trust those locations. By fixing the reference points throughout the decoding layers, the model can ignore the "where" and focus entirely on the "what"—the semantic context surrounding the road elements.
  • Relational Graph Attention Network (RGAT): Road elements are not independent. The RGAT captures inter-instance dependencies (e.g., how one lane divider relates to its neighbor) to ensure the predicted map is globally consistent.
  • Unified Framework: As shown in Figure 2, the model uses a shared backbone and encoder, with task-specific heads that generate predictions for intersection topology, merging/diverging points, or drivable area boundaries.
Figure 2: Framework of the PGU-Decoder.

Datasets

Because no public benchmark covers the specific road-level and lane-level data required for industrial updating, the authors constructed three task-specific datasets using AutoNavi Map (Amap) data:

  • Cities: Data from 12 major Chinese cities (e.g., Beijing, Guangzhou, Shanghai). 8 cities were used for training/validation, and 4 were held out for testing to evaluate generalization.
  • Scale:
    • Intra-intersection: 125,929 intersection samples.
  • Merging/Diverging: 25,720 lane divider pairs.
  • Drivable-Area: 42,716 intersection samples.
  • Production Validation: The model was eventually deployed across more than 360 cities in China.

Key Insights and Ideas

  • Priors are Power: In industrial settings, you rarely start from a blank slate. Using existing coarse map data as a "hard prior" is significantly more effective than letting a model try to localize elements from scratch using raw sensor data.
  • Stability over Refinement: The comparison in Figure 3 highlights a counter-intuitive finding: for map updating, not allowing the model to refine its search coordinates (fixed reference points) leads to better results. This stability prevents the model from "drifting" away from reliable upstream geometries.
  • Relational Context Matters: Drivable-area detection and topology generation are highly dependent on the relationship between neighboring road segments. The inclusion of RGAT was crucial for reaching production-level accuracy.
Figure 3: Comparison of reference-point mechanisms.

Key Results

PGU-Decoder was compared against seven baselines, including specialized models like YOLOv11 and DINO, and even large vision-language models like Qwen3-VL.

  • Accuracy: It consistently outperformed all baselines. For example, it achieved 98.12% Intersection Accuracy (I-Acc), surpassing a Graph Attention Network (GAT) baseline by over 5%.
  • Industrial Impact:
    • Quality: Production quality scores exceeded 98% across all three tasks.
  • Cost: Reduced manual verification rates from 100% to as low as 3%, resulting in a 95% reduction in manual processing costs (saving an estimated CNY 39 million annually).
  • Speed: Improved map update frequency from weekly to hourly.

Limitations and Future Work

  • Upstream Dependency: The framework is heavily dependent on the quality of upstream geometries. If the initial coarse map is missing a road entirely, the decoder might not have a prior query to start with.
  • Uncertainty Handling: The authors suggest that future work should investigate "uncertainty-aware decoding" to better handle cases where upstream data is known to be erroneous or incomplete.
  • Cross-Source Transfer: Evaluating how well the model transfers across different map sources or production specifications in different countries remains an open area for exploration.

Summary updated

Abstract

Timely road-network map updates are essential to navigation and autonomous driving. Three key tasks are intra-intersection road network generation, merging/diverging detection, and drivable-area detection, covering topology, geometry, and attributes, respectively. Separate pipelines duplicate rule maintenance, quality assurance, and manual verification, increasing operational costs. We propose PGU-Decoder, a production-grade unified decoding framework for semantic prediction across these tasks. A Prior Query Module (PQM) constructs point-level prior queries from sampled locations on road-network geometries. Fixed reference points preserve positional priors throughout decoding, allowing the model to focus on surrounding context. Point-level representations are aggregated into instance-level representations, and a Relational Graph Attention Network (RGAT) captures inter-instance dependencies. Task-specific heads then generate predictions. We validate PGU-Decoder in large-scale automated road-network map updating across more than 360 cities in China, where it consistently outperforms seven baselines. Following deployment at AutoNavi Map, it achieves quality scores above 98% on all three tasks, reduces manual verification costs by 95%, and increases update frequency from weekly to hourly. These results demonstrate a practical and cost-effective industrial solution for road-network map updating.
BibTeX
@misc{ning2026pgu,
  author = {Yuguang Ning and Bingbing Kang and Nan Zhao and Zhixuan Li and Ze Zhao and Dongdong Liu and Wenjian Zhu and Zekai Liu and Junhui Wang},
  title = {{PGU-Decoder: Prior-Guided Unified Decoding for Industrial Road-Network Map Updating}},
  year = {2026},
  eprint = {2611.23936},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23936},
}

Loading related content...

Figure