RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction

Chenxu Peng, Chenxu Wang, Yimian Dai, Yongxiang Liu, Ming-Ming Cheng, Xiang Li

TL;DR

Extracting accurate road networks from aerial imagery is notoriously difficult due to complex topologies and occlusions. This paper introduces WorldRoadSeg-360K, the largest and most geographically diverse road segmentation dataset to date, alongside RoadGIE, an ultra-lightweight (3.7M parameters) interactive framework. By leveraging connectivity-aware prompts like scribbles and a novel prompt-excluded loss, RoadGIE achieves state-of-the-art topological consistency while slashing manual annotation time by 79%.

What Problem Does This Paper Address?

Fully automated road extraction from remote sensing imagery frequently fails because roads are thin, elongated, and easily obscured by trees, buildings, or shadows. This leads to fragmented, disconnected road maps that are useless for real-world navigation systems.

To fix these errors, humans need to be in the loop. However, existing interactive segmentation models (like the Segment Anything Model, or SAM) rely on point or bounding box prompts. These prompt types are poorly suited for capturing the continuous, linear topology of road networks. Furthermore, existing interactive frameworks suffer from interactive degradation—as a user adds more corrective prompts over multiple rounds, the model can actually lose track of previously correct structures. Finally, progress has been bottlenecked by a lack of diverse, high-quality training data; existing datasets are typically confined to a few cities or lack pixel-level width annotations.

Methods

The authors present RoadGIE, a real-time interactive road segmentation framework designed specifically around road morphology.

Figure 3. Overview of our RoadGIE pipeline. Given an input image and an initial guidance prompt, the network produces a preliminary segmentation. This output is compared against the ground truth to compute training loss and localize misclassified regions. A simulated annotator then generates corrective hints in response to these errors. These updated prompts, combined with the model’s latest prediction, are fed back into the network for iterative refinement.

Key innovations include:

  • Directional Aggregation Module (DAM): Built on top of a lightweight UNet decoder, the DAM uses efficient 1D convolutions along four dominant directions (vertical, horizontal, diagonal, and anti-diagonal) to capture long-range linear dependencies and preserve structural continuity.
  • Connectivity-Aware Prompt Simulation: To train the model without manual prompt data, the authors simulate realistic user interactions, generating not just clicks but also center, line, and Bezier curve scribbles within misclassified regions.
  • Topo-Semantic Coupled Road Instantiation: To resolve user intent ambiguity (e.g., whether a user wants to select just a highway or all connecting local roads), the model extracts centerlines and groups segments before generating the final mask.
  • Prompt-Excluded Skeleton Loss: To combat interactive degradation, the authors adapted the skeleton-based recall loss. Crucially, they mask out the prompt areas, applying this structural loss only to non-prompt regions. This prevents the model from overfitting to localized user inputs and forces it to maintain global geometric consistency.

Datasets

The paper introduces WorldRoadSeg-360K, a massive leap forward in dataset scale and diversity for this domain.

Composition, construction workflow, and global distribution of the WorldRoadSeg-360K dataset.
  • WorldRoadSeg-360K (New): Contains 366,947 high-resolution (0.8–1.1m) satellite images and masks across 223 cities in 38 countries, spanning diverse terrains and environments. It was built by using coarse OpenStreetMap data to prompt foundation models (SAM, HQ-SAM, RobustSAM), followed by rigorous manual validation.
  • LSRV Dataset: Used exclusively as an out-of-domain (OOD) test set to evaluate cross-domain generalization.
  • Baseline Dataset: A combination of 10 older road datasets (e.g., SpaceNet, DeepGlobe, LoveDA) used for benchmarking.

Key Insights and Ideas

  • Scribbles Over Points: For high-aspect-ratio structures like roads, sparse point clicks provide weak topological cues. Simulating and utilizing continuous Bezier scribbles allows the model to capture connectivity priors much more effectively.
  • The Power of Prompt Exclusion: Standard loss functions can cause a model to over-optimize on the exact pixels a user clicked, leading to a degradation of the surrounding road structure in later interaction rounds. Restricting the skeleton recall loss to unprompted regions acts as a powerful regularizer.
Visual comparison demonstrating how Prompt-Excluded Skeleton Loss preserves thin road structures across multiple interaction steps.

Key Results

RoadGIE sets a new state-of-the-art across multiple benchmarks, outperforming heavyweights like ScribblePrompt and SAM while remaining incredibly compact.

  • Segmentation Performance: On the WorldRoadSeg-360K dataset, RoadGIE achieved a Dice score of 0.835 and an APLS (Average Path Length Similarity) of 0.620, outperforming the closest competitor (ScribblePrompt) by 2.6 Dice points.
  • Efficiency: The model contains only 3.7M parameters and boasts a GPU inference runtime of just 39.52 ms, making it highly viable for real-time, low-resource deployment.
  • User Study: In a practical test annotating 100 images, using RoadGIE improved annotation accuracy (Dice increased from 0.827 to 0.885) while reducing annotation time by 79% (from 73 seconds to just 15 seconds per image).

Limitations and Future Work

While highly effective, the authors note a few limitations:

  1. Resolution Constraints: The dataset and model are optimized for 0.8–1.1m spatial resolution imagery. Performance might vary when applied to ultra-high-resolution data.
  2. Interaction Step Limits during Training: Due to GPU memory constraints, the model was trained using a maximum of 6 interaction steps. In highly complex real-world scenes, users might require more iterations, which could slightly degrade accuracy. Future work will look into scaling the interaction depth during training.

Summary updated

Abstract

Accurate road segmentation from aerial imagery is fundamental to many geospatial applications. However, existing datasets often suffer from limited scene diversity, low semantic granularity, and poor structural continuity, restricting their generalization across environments. To address these challenges, we introduce WorldRoadSeg-360K, the largest and most diverse road segmentation dataset to date, comprising 366,947 high-resolution images collected from 38 countries and 223 cities across various terrains and continents. WorldRoadSeg-360K serves as a comprehensive benchmark and reveals key challenges in handling diverse and structurally complex scenes. Automated approaches often struggle to preserve road connectivity, while current interactive methods lack efficient, topology-sensitive tools for real-world road editing. To this end, we present RoadGIE, establishing a novel interactive paradigm for road extraction in remote sensing. Unlike prior point- or box-based prompting strategies, RoadGIE supports connectivity-aware prompts, including clicks and scribbles, which inherently align with the topology of road networks. To improve structural consistency and mitigate performance degradation during iterative interactions, RoadGIE integrates an expert-guided prompting strategy and adapts the skeleton-based recall loss for interactive scenarios. RoadGIE achieves state-of-the-art performance in both segmentation accuracy and topological consistency on WorldRoadSeg-360K and other benchmarks, while maintaining efficient operation with only 3.7M parameters. The code are publicly available at: https://github.com/chaineypung/RoadGIE
BibTeX
@inproceedings{peng2026roadgie,
  author = {Chenxu Peng and Chenxu Wang and Yimian Dai and Yongxiang Liu and Ming-Ming Cheng and Xiang Li},
  title = {{RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=xdayvR3tbU},
}
Figure