Beyond Endpoints: Path-Centric Reasoning for Vectorized Off-Road Network Extraction

Wenfei Guan, Jilin Mei, Tong Shen, Xumin Wu, Shuo Wang, Min Cheng, Yu Hu

Extracting road networks from satellite imagery is a cornerstone of modern navigation, but most current models have a "city bias." They excel at structured, high-contrast urban grids but crumble when faced with the messy reality of off-road terrain—dirt tracks obscured by tree canopies, faint trails in deserts, and ambiguous junctions in the wild.

A new paper from researchers at the Chinese Academy of Sciences, titled "Beyond Endpoints: Path-Centric Reasoning for Vectorized Off-Road Network Extraction," introduces a novel framework and dataset designed to bridge this "urban-to-off-road" domain gap.

TL;DR

The authors move beyond "node-centric" models that only look at road junctions, introducing MaGRoad, a framework that reasons about the entire path between points to extract road networks in challenging wild terrains. To power this, they released WildRoad, the first large-scale, continent-spanning vectorized dataset for off-road environments, along with an AI-assisted tool that makes annotation significantly faster.

What Problem Does This Paper Address?

Most state-of-the-art road extraction models (like SAM-Road) are node-centric. They identify potential junctions (nodes) and then try to guess if they are connected based primarily on the visual features at those specific endpoints.

In a city, this works fine. But in the wild, two endpoints might look like they belong to a road, yet the space between them is a dense forest or a cliff. Conversely, a road might be perfectly visible in the middle but obscured by shadows at the junction. Node-centric models are "fragile" to these local ambiguities, leading to fragmented maps and incorrect topologies. Furthermore, there has been a historical lack of high-quality, vectorized (graph-based) datasets for off-road scenes to train better models.

Methods

The authors propose MaGRoad (Mask-aware Geodesic Road network extractor), which shifts the paradigm from node-centric to path-centric reasoning.

  1. MaGTopoNet: Instead of just looking at the start and end of a potential road segment, this module samples "evidence" (visual features) at 32 points along the entire geodesic path between them. It calculates statistics like the mean, standard deviation, and a "softmin" (to detect bottlenecks or breaks) of the road probability along that path.
  2. Edge-Biased Attention: The model uses a self-attention mechanism that forces candidate edges to "compete." This helps the model choose the most plausible connection while suppressing spurious ones.
  3. Efficient Vertex Extraction: Standard pipelines use three separate passes of Non-Maximum Suppression (NMS) to find road points, which is slow. MaGRoad unifies this into a single, refactored NMS pass, achieving a significant speedup.
  4. Interactive Annotation Pipeline: To build their dataset, they created a "Prompt-Propose-Refine" workflow. A human provides a few sparse clicks, the AI proposes a graph, and the human quickly corrects it. This is much faster than manual tracing.

Datasets

  • WildRoad (New): The star of the paper. It contains 221 high-resolution images (8K x 4K) covering 2,100 $km^2$ across six continents. Unlike previous rural datasets that only provide pixel masks, WildRoad provides full vectorized graphs.
  • Urban Benchmarks: To test generalization, they evaluated on City-Scale, SpaceNet, and Global-Scale.

Key Insights and Ideas

  • The Path is the Proof: The core "aha" moment is that the connectivity of a road is a global property, not a local one. By aggregating multi-scale visual evidence along the entire candidate path, the model can "see through" a tree covering a junction if the rest of the path is clear.
  • Geometric Competition: By using an additive bias in the attention mechanism, the model learns that if one path is very strong, nearby overlapping paths are likely incorrect. This mimics how humans perceive distinct road segments.
  • Unified NMS: The researchers found that the computational bottleneck wasn't just the neural network, but the post-processing. Simplifying the vertex extraction logic provided a massive practical boost for large-scale mapping.

Key Results

  • State-of-the-Art Performance: On the WildRoad benchmark, MaGRoad outperformed the previous best model (SAM-Road++) in both graph completeness (F1 score) and topological accuracy (APLS).
  • Efficiency: The "MaGRoad-fast" variant achieved a 2.5× speedup in inference time (27.8 minutes vs. 76.1 minutes for a standard benchmark run) while actually improving the F1 score.
  • Robust Generalization: Despite being designed for the "wild," MaGRoad remained highly competitive on urban datasets, proving that path-centric reasoning is a fundamentally better way to think about roads, regardless of the environment.

Limitations and Future Work

  • Topological Noise: While the "fast" version of the model has higher recall (it finds more roads), it introduces slightly more topological noise, leading to a small drop in the APLS metric compared to the base model.
  • Imagery Dependence: The model still relies on high-resolution RGB imagery. Performance in extremely dense, multi-story jungle canopies where the ground is 100% obscured remains a challenge for any optical-based system.
  • Future Directions: The authors suggest that integrating this path-centric approach with other data sources (like GPS traces or LiDAR) could further improve robustness in the most extreme environments.

Check out the project: The authors have released the WildRoad dataset and the MaGRoad code at https://github.com/xiaofei-guan/MaGRoad.

Summary updated

Abstract

Deep learning has advanced vectorized road extraction in urban settings, yet off-road environments remain underexplored and challenging. A significant domain gap causes advanced models to fail in wild terrains due to two key issues: lack of large-scale vectorized datasets and structural weakness in prevailing methods. Models such as SAM-Road employ a node-centric paradigm that reasons at sparse endpoints, making them fragile to occlusions and ambiguous junctions in off-road scenes, leading to topological errors. This work addresses these limitations in two complementary ways. First, we release WildRoad, a global off-road road network dataset constructed efficiently with a dedicated interactive annotation tool tailored for road-network labeling. Second, we introduce MaGRoad (Mask-aware Geodesic Road network extractor), a path-centric framework that aggregates multi-scale visual evidence along candidate paths to infer connectivity robustly. Extensive experiments show that MaGRoad achieves state-of-the-art performance on our challenging WildRoad benchmark while generalizing well to urban datasets. An efficient vertex extraction strategy also yields roughly 2.5X faster inference, improving practical applicability. Together, the dataset and path-centric paradigm provide a stronger foundation for mapping roads in the wild. We release both the dataset and code at this repository. We release both the dataset and code at https://github.com/xiaofei-guan/MaGRoad.
BibTeX
@misc{guan2025beyond,
  author = {Wenfei Guan and Jilin Mei and Tong Shen and Xumin Wu and Shuo Wang and Min Cheng and Yu Hu},
  title = {{Beyond Endpoints: Path-Centric Reasoning for Vectorized Off-Road Network Extraction}},
  year = {2025},
  eprint = {2512.10416},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.10416},
}
Figure