💻 computer science

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

This paper introduces RoadGIE, a novel interactive road extraction framework that leverages connectivity-aware prompts and an expert-guided strategy to achieve state-of-the-art performance on the newly proposed WorldRoadSeg-360K, the largest and most diverse global-scale aerial road segmentation dataset to date.

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

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a perfect map of every single road in the world using only satellite photos. It's a massive, messy job. Roads twist, turn, get hidden by trees, disappear under bridges, and look very different in a busy city compared to a quiet desert.

This paper introduces two main things to solve this problem: a giant new library of road photos and a smart, interactive drawing assistant that learns how to draw roads better than any computer program has before.

Here is the breakdown in simple terms:

1. The Problem: The Old Maps Were Broken

Previously, computer programs trying to find roads had two big headaches:

Not enough variety: They were mostly trained on photos from just one or two cities. If you showed them a road in a different country or a different type of terrain (like mountains), they got confused.
Bad at "connectivity": Roads are long, connected lines. Old programs often treated them like a string of disconnected beads. They would draw a road, stop, leave a gap, and start again, making the map useless for navigation.
Hard to fix: If a computer made a mistake, fixing it was slow and frustrating.

2. The Solution Part A: "WorldRoadSeg-360K" (The Giant Library)

The authors built the biggest, most diverse road dataset ever created.

The Scale: Imagine a photo album with 366,947 high-resolution pictures.
The Reach: These photos come from 38 different countries and 223 cities, covering almost every continent (except Antarctica).
The Variety: It includes everything from dense city centers to rural villages and winding mountain passes.
The Quality: They didn't just copy existing maps. They used a "smart fusion" process: they took rough map data, ran it through several advanced AI models, and then had humans check the results to make sure the road edges were perfect.

Analogy: Think of this dataset as a master chef gathering ingredients from every corner of the globe to create the ultimate recipe book. Before, chefs only had recipes for one type of soup; now, they have a library that teaches them how to cook soup in any climate.

3. The Solution Part B: "RoadGIE" (The Smart Drawing Assistant)

Even with a great library, computers still struggle to draw roads perfectly on their own. So, the authors created RoadGIE, a tool that lets a human and a computer work together in real-time.

How it works differently:

Old Way: You had to click a single dot or draw a box around a road. The computer guessed the rest. This often failed because a dot doesn't tell the computer if the road curves left or right.
RoadGIE Way: You can click or scribble (draw a quick line) along the road.
- The Magic: The system understands that a scribble isn't just a line; it's a connection. It knows roads are continuous paths. If you scribble over a bridge, it knows the road continues underneath, even if it's hidden.

Key Features of RoadGIE:

The "Expert Guide": Sometimes the computer gets stuck on tricky spots (like a road hidden by shadows). RoadGIE has a built-in "expert" that tells the computer, "Hey, look here, this part is confusing, pay extra attention." This helps the computer learn faster.
The "Skeleton Loss": Imagine you are drawing a road, and you accidentally erase a tiny piece of it. Most programs would just leave it blank. RoadGIE has a special rule (a mathematical "loss function") that says, "Don't forget the skeleton! Keep the road connected even if you make a mistake." It ensures the road stays as one long, unbroken line.
Speed: It is incredibly fast and lightweight (only 3.7 million parameters), meaning it can run on standard computers without needing a supercomputer.

4. The Results: Faster and Better

The authors tested RoadGIE against other top programs and found:

Accuracy: It draws roads more accurately than any previous method, especially when dealing with complex, winding roads.
Speed: It cuts the time needed to map a road by 79%. Instead of a human spending over a minute to draw a road manually, RoadGIE helps them do it in about 15 seconds with high precision.
Generalization: Because it was trained on the giant "WorldRoadSeg-360K" library, it works well on roads it has never seen before, whether in the USA, China, or Africa.

Summary

Think of this paper as introducing a new GPS system for map-makers.

They built a massive, global training school (the dataset) so the AI learns what roads look like everywhere.
They built a smart assistant (RoadGIE) that understands how humans draw roads (using scribbles) and knows that roads must stay connected, even when they are hard to see.

The result is a tool that makes creating accurate road maps faster, cheaper, and more reliable than ever before.

Technical Summary: RoadGIE and WorldRoadSeg-360K

Problem Statement

Accurate road segmentation from aerial imagery is critical for geospatial applications such as road modeling, map updating, and spatial structure analysis. However, existing approaches face three primary challenges:

Data Limitations: Current datasets often lack scene diversity, semantic granularity, and structural continuity. Many are restricted to single cities or regions, while global datasets often rely on vectorized centerlines that miss road width and boundary continuity, hindering fine-grained segmentation.
Model Generalization: Fully automatic segmentation models frequently produce fragmented road maps, failing to preserve the topological connectivity required for high-precision modeling.
Interactive Inefficiency: While interactive models (e.g., the SAM family) show promise, their standard point- or box-based prompting strategies are ill-suited for the elongated, continuous nature of roads. These methods often struggle with ambiguous boundaries, occlusions, and user intent ambiguity, leading to high latency and degraded performance during iterative refinement.

Methodology

The paper proposes a dual contribution: a new large-scale dataset (WorldRoadSeg-360K) and a novel interactive framework (RoadGIE).

1. WorldRoadSeg-360K Dataset

To address data scarcity and diversity issues, the authors constructed a global-scale benchmark:

Scale & Diversity: Comprises 366,947 high-resolution images (512×512 pixels, 0.8–1.1m resolution) collected from 38 countries and 223 cities across all continents except Antarctica.
Annotation Quality: Annotations were generated by using OpenStreetMap (OSM) road maps as prompts for state-of-the-art segmentation models (SAM, HQ-SAM, RobustSAM). The outputs were fused and manually validated to produce pixel-level masks.
Evaluation: Includes an Out-of-Domain (OOD) test set derived from the LSRV dataset to rigorously assess cross-domain generalization.

2. RoadGIE Framework

RoadGIE is a real-time, interactive road extraction framework designed to handle the topological sensitivity of road networks. It operates iteratively, refining predictions based on user or simulated prompts.

Network Architecture:
- Utilizes a lightweight UNet backbone (3.7M parameters) for efficiency.
- Introduces a Directional Aggregation Module (DAM) at the decoder. This module applies 1D convolutions along four directions (vertical, horizontal, diagonal, anti-diagonal) to capture long-range dependencies and enforce directional consistency, crucial for preserving road connectivity.
Connectivity-Aware Prompting:
- Unlike point/box prompts, RoadGIE supports clicks and scribbles.
- Prompt Simulation: During training, the system simulates user interactions by identifying prediction errors and generating synthetic corrections. These include center scribbles (skeleton-based), line scribbles (straight strokes), and Bezier scribbles (smooth curves), which better encode the shape and continuity of roads.
- Expert-Guided Prompting: A strategy is employed to simulate prompts in high-uncertainty regions (where multiple models disagree), guiding the model to focus on difficult areas like occluded or blurred segments.
Resolving User Ambiguity:
- To handle varying user intents (e.g., selecting a specific highway vs. all roads), the framework employs Topo-Semantic Coupled Road Instantiation.
- This process first regularizes road structure, extracts centerlines, and groups segments. A learnable scorer then ranks candidates based on the input prompt, and the top segment is iteratively expanded and refined. This defers final instantiation until structural abstraction is complete, improving generalization.
Loss Function:
- To prevent performance degradation during iterative interactions (where later prompts might overwrite correct regions), the authors introduce a Prompt-Excluded Skeleton Loss.
- This loss applies skeleton-based supervision (specifically recall loss) only to non-prompt regions. This encourages the model to learn structural consistency in uncertain areas without overfitting to the already well-supervised prompt regions.

Key Contributions

WorldRoadSeg-360K: The largest and most diverse road segmentation dataset to date, providing a universal benchmark for evaluating model performance across varied terrains and geographies.
RoadGIE Framework: The first generalizable interactive framework for road extraction that supports connectivity-aware prompts (clicks and scribbles). It achieves state-of-the-art (SOTA) accuracy and topological consistency while maintaining a compact model size (3.7M parameters).
Comprehensive Interaction Strategy: A novel combination of expert-guided prompting, topo-semantic instantiation, and prompt-excluded skeleton loss to reduce user intent ambiguity and maintain structural consistency across interaction rounds.

Experimental Results

Performance: RoadGIE outperforms existing SOTA models (including SAM variants, ScribblePrompt, and EISeg) on both the WorldRoadSeg-360K and LSRV benchmarks.
- On WorldRoadSeg-360K, it achieves a Dice score of 0.835 and APLS of 0.620, surpassing the second-best model (ScribblePrompt) by 2.6 Dice points.
- It demonstrates superior performance in topological metrics (clDice, $\beta_0$ , $\beta_1$ ), indicating better connectivity and fewer fragmented components.
Ablation Studies:
- Training on WorldRoadSeg-360K significantly improves downstream performance compared to training on smaller or less diverse datasets.
- Bezier scribbles yielded the best performance among prompting styles, confirming the value of fine-grained structural cues.
- The Expert-Guided Prompt improved the Dice score by 2.7% and APLS by 2.9% at the 5th interaction step compared to the baseline.
- The Prompt-Excluded Skeleton Loss was essential for maintaining structural integrity, outperforming standard Focal and Dice losses in preserving road connectivity.
Efficiency:
- Annotation Speed: RoadGIE reduces annotation time by 79% compared to manual labeling (15s vs. 73s per image) while achieving expert-level accuracy (Dice 0.885 vs. 0.827).
- Inference: The model runs at 39.52 ms per prediction on an NVIDIA RTX 3090, making it suitable for real-time applications.

Significance and Claims

The paper claims that WorldRoadSeg-360K serves as a comprehensive resource that reveals the limitations of current automated and interactive methods in handling structurally complex scenes. By introducing RoadGIE, the authors establish a new paradigm for remote sensing road extraction that aligns visual prompting strategies with the morphological characteristics of road networks.

The work demonstrates that by integrating connectivity-aware prompts (scribbles) with topology-sensitive training objectives (skeleton loss) and structural abstraction (topo-semantic instantiation), it is possible to achieve high-precision, topologically consistent road segmentation with minimal user effort. The authors position RoadGIE as a cost-effective, real-time solution that bridges the gap between fully automatic methods (which lack precision) and traditional manual annotation (which is too slow).

Limitations noted by the authors include the dataset's resolution range (0.8–1.1m), which may limit applicability to higher-resolution scenarios, and the constraint of training on six interaction steps due to GPU memory limits, which may affect performance in extremely complex scenes requiring more iterations.