WorldRoadSeg-360K: Global Road Segmentation Benchmark

• WorldRoadSeg-360K is a globally diverse, single-class road segmentation benchmark comprising 366,947 512×512 aerial image patches with binary masks.
• It addresses prior limitations by enhancing scene diversity, semantic granularity, and structural continuity, supporting both automatic and interactive extraction.
• The dataset uses a semi-automatic pipeline that refines OSM-derived annotations via SAM-family models, enabling topology-sensitive evaluation with metrics like Dice, clDice, and APLS.

Searching arXiv for the primary paper and closely related road-segmentation benchmark work. WorldRoadSeg-360K is a large-scale, pixel-level aerial road segmentation dataset introduced in “RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction” (Peng et al., 26 May 2026). It comprises 366,947 high-resolution images, each of size $512 \times 512$, with spatial resolution $0.8\text{–}1.1$ m per pixel, collected from 38 countries and 223 cities across all continents except Antarctica (Peng et al., 26 May 2026). The benchmark is designed to address three limitations identified in prior road datasets—limited scene diversity, low semantic granularity, and poor structural continuity—and to support both automatic road segmentation and interactive road extraction under topology-sensitive evaluation (Peng et al., 26 May 2026).

1. Definition, scope, and benchmark role

WorldRoadSeg-360K is a single-class road segmentation benchmark with pixel-wise binary masks, i.e., road versus background, rather than graph centerlines or multi-class road taxonomies (Peng et al., 26 May 2026). Its scene distribution includes urban and rural settings, explicitly covering dense urban centers, rural outskirts, and mountainous areas (Peng et al., 26 May 2026). The dataset is presented as the largest and most geographically diverse pixel-level road segmentation dataset at the time of publication, and as a universal benchmark for road segmentation and interactive road extraction in remote sensing (Peng et al., 26 May 2026).

The benchmark is motivated by the observation that many earlier datasets are geographically narrow, often centered on a small number of cities or a single country, and that line-based annotations omit road width and boundary continuity (Peng et al., 26 May 2026). In this framing, WorldRoadSeg-360K is not only a corpus for per-pixel prediction but also a pretraining resource intended to improve downstream performance on other road datasets such as LSRV (Peng et al., 26 May 2026). The benchmark further emphasizes failure modes that are weakly reflected by region-overlap metrics alone, including occlusions, elongated roads, and complex intersections, thereby making topology-aware evaluation central rather than auxiliary (Peng et al., 26 May 2026).

A common misconception is to treat WorldRoadSeg-360K as a generic remote-sensing segmentation dataset of “360K items.” In fact, its 366,947 samples are $512 \times 512$ aerial image patches with binary road masks, and its scope is road extraction rather than generic land-cover change analysis (Peng et al., 26 May 2026). This distinguishes it from QAG-360K, a separate benchmark for Change Detection Question Answering and Grounding built around question–answer–mask triplets rather than road segmentation images (Li et al., 2024).

2. Geographic scale, patching scheme, and comparison to earlier datasets

WorldRoadSeg-360K is explicitly constructed at global scale, with imagery from 223 cities in 38 countries and coverage across all continents except Antarctica (Peng et al., 26 May 2026). The paper states that its geographic coverage is roughly four times larger than the Global-Scale road graph dataset, while also providing mask annotations rather than line or graph representations (Peng et al., 26 May 2026). This combination of global diversity and mask-based supervision is central to its positioning as a benchmark for generalizable road extraction.

For fair comparisons with existing datasets, all datasets in the benchmark comparison are uniformly divided into $512 \times 512$ patches, and patches with no roads or mostly black backgrounds are removed (Peng et al., 26 May 2026). The paper does not specify file formats or georeferencing fields, although the textual description implies city-level and country-level organization (Peng et al., 26 May 2026).

Dataset	Images (512×512 patches)	Annotation
Massachusetts	10,413	Line
DeepGlobe	24,904	Mask
LSRV	1,787	Mask
Global-Scale	55,488	Line
WorldRoadSeg-360K	366,947	Mask

Within this comparison, WorldRoadSeg-360K differs from graph-labeled datasets such as Massachusetts, SpaceNet, and Global-Scale by supplying pixel-level masks that preserve road width and boundary continuity (Peng et al., 26 May 2026). It also differs from smaller mask-based datasets such as WHU-road, HUAWEI, LoveDA, and LSRV by substantially expanding city count and scene diversity (Peng et al., 26 May 2026). This suggests that the benchmark is intended not merely to increase sample count, but to change the statistical structure of training and evaluation toward cross-region robustness.

3. Data collection and annotation pipeline

The collection process begins with a global urban area survey in which rectangular regions of size 15–45 km are selected to represent cities of different scales and terrain types, including dense metropolitan cores, surrounding rural areas, and mountainous or complex terrain (Peng et al., 26 May 2026). High-resolution satellite imagery is then downloaded via the Google Static Maps API, with the dataset following the $0.8\text{–}1.1$ m resolution standard of Sat2Graph (Peng et al., 26 May 2026). The paper does not specify sensor vendors, bands, or acquisition dates (Peng et al., 26 May 2026).

Initial labels are obtained from OpenStreetMap road centerlines, used as coarse graph-like seeds rather than final supervision (Peng et al., 26 May 2026). These OSM annotations are then used to prompt several segmentation models—SAM, HQ-SAM, and RobustSAM—to infer full road masks (Peng et al., 26 May 2026). The outputs from these models are fused and combined with the original annotations to produce refined road masks, followed by a manual validation phase in which samples are categorized into high- and low-quality subsets (Peng et al., 26 May 2026). High-quality samples are intended for further fine-tuning during model training (Peng et al., 26 May 2026).

This annotation design has two notable implications. First, the benchmark is mask-based despite beginning from centerline supervision, which directly addresses the width and continuity limitations of line-only datasets (Peng et al., 26 May 2026). Second, the semi-automatic pipeline implies that annotation quality depends partly on OSM completeness and the behavior of SAM-family models, a point the paper later reflects in its discussion of label noise and quality partitioning (Peng et al., 26 May 2026).

The paper does not provide pixel-level labeling rules such as the treatment of sidewalks, parking lanes, or minimum mapped road width (Peng et al., 26 May 2026). For an encyclopedia treatment, that omission is important: WorldRoadSeg-360K is richly specified in scale and benchmark intent, but under-specified in some annotation semantics.

4. Tasks, evaluation metrics, and protocol design

WorldRoadSeg-360K supports two conceptual tasks: automatic road segmentation and interactive road segmentation or editing (Peng et al., 26 May 2026). The latter is particularly emphasized because the benchmark is introduced together with RoadGIE, an interactive framework designed for road morphology and connectivity-aware prompting (Peng et al., 26 May 2026).

The evaluation protocol adopts six metrics: Dice coefficient, Recall, clDice, APLS, $\beta_0$, and $\beta_1$ (Peng et al., 26 May 2026). Dice measures region overlap, Recall measures sensitivity of road detection, clDice is a connectivity-preserving Dice metric for tubular structures, APLS evaluates shortest-path similarity between predicted and ground-truth graphs, and the Betti numbers $\beta_0$ and $\beta_1$ quantify connected components and loops, respectively (Peng et al., 26 May 2026). The inclusion of topological metrics is not incidental: the benchmark is explicitly designed around structural continuity and connectivity, which are often weakly captured by overlap metrics alone (Peng et al., 26 May 2026).

The standard Dice expression is given as

$$\text{Dice} = \frac{2 |Y \cap \hat{Y}|}{|Y| + |\hat{Y}|},$$

with $0.8\text{–}1.1$0 the ground-truth mask and $0.8\text{–}1.1$1 the prediction (Peng et al., 26 May 2026). For training RoadGIE, the paper uses a composite objective consisting of focal loss, soft Dice loss, and a prompt-excluded skeleton-based recall loss adapted from prior skeleton-based supervision (Peng et al., 26 May 2026). The skeleton term is restricted to non-prompt pixels, so the model is encouraged to preserve connectivity even where user guidance is sparse (Peng et al., 26 May 2026).

During training, each batch runs 5 interaction rounds with 1–3 visual prompts per round (Peng et al., 26 May 2026). For evaluation, metrics are typically computed after 5 interaction steps on the LSRV test set, although the paper also analyzes performance up to 10 iterations and across different prompt types (Peng et al., 26 May 2026). This means that WorldRoadSeg-360K is not only a static image-mask dataset; it is embedded in a benchmark protocol that treats iterative correction as part of the task formulation.

5. RoadGIE and the interactive road extraction paradigm

RoadGIE is the model introduced alongside WorldRoadSeg-360K and serves as the principal benchmark method (Peng et al., 26 May 2026). It is a lightweight UNet-based interactive model with a Directional Aggregation Module, designed for connectivity-aware decoding, and has 3.7M parameters with inference time of approximately 39.52 ms per $0.8\text{–}1.1$2 image on RTX3090 (Peng et al., 26 May 2026). At interaction step $0.8\text{–}1.1$3, the model takes the image $0.8\text{–}1.1$4, previous prediction mask $0.8\text{–}1.1$5, and current prompt map $0.8\text{–}1.1$6, and outputs an updated prediction

$0.8\text{–}1.1$7

(Peng et al., 26 May 2026)

Its distinctive feature is support for connectivity-aware prompts, including clicks and scribbles (Peng et al., 26 May 2026). The scribble types include center scribbles derived from the skeleton, straight line scribbles within error regions, and Bézier scribbles through sampled control points (Peng et al., 26 May 2026). The paper reports that Bézier scribbles achieve the best performance, while point prompts lag, which is interpreted as evidence that connectivity-aware prompts align better with high-aspect-ratio road networks (Peng et al., 26 May 2026).

RoadGIE also introduces expert-guided prompting. Given predictions from pre-trained segmentation models, an uncertainty map

$0.8\text{–}1.1$8

is used to bias prompt placement toward hard regions such as occluded roads (Peng et al., 26 May 2026). After 5 rounds, this strategy improves Dice from 84.9 to 87.6 and APLS from 63.7 to 66.6 in the paper’s per-round analysis (Peng et al., 26 May 2026).

A further component, termed topo-semantic road instantiation in the paper, refines road structure, extracts centerlines, computes segment attributes, groups segments into candidates, and uses prompt-conditioned selection and iterative expansion to disambiguate which road segments correspond to a prompt (Peng et al., 26 May 2026). This suggests an effort to move beyond generic interactive segmentation toward road-specific structural reasoning.

The broader significance is that WorldRoadSeg-360K is inseparable from an interactive view of road extraction. Earlier fast road-segmentation systems such as LFD-RoadSeg are optimized for dense road versus non-road prediction on on-road imagery, emphasizing low-level features, bilateral structure, and runtime efficiency (Zhou et al., 2024). By contrast, RoadGIE targets aerial imagery, topology-sensitive prompting, and interactive correction at global scale (Peng et al., 26 May 2026). The two efforts are related by task family but operate under different data regimes and user assumptions.

6. Empirical results, generalization, and limitations

The paper reports that training on WorldRoadSeg-360K improves the performance of all tested methods relative to training on a baseline dataset composed of older road segmentation datasets (Peng et al., 26 May 2026). On the common LSRV test set after 5 interaction steps, RoadGIE achieves Dice 0.835 and APLS 0.620 when trained on WorldRoadSeg-360K, compared with Dice 0.807 and APLS 0.593 when trained on the baseline dataset (Peng et al., 26 May 2026). Other interactive baselines such as EISeg, ScribbleSeg, SAM, PRISM-2D, PRISM-2D-Lite, and ScribblePrompt also improve when trained on WorldRoadSeg-360K, though RoadGIE remains best in both settings (Peng et al., 26 May 2026).

Cross-dataset pretraining comparisons reinforce this point. When pretrained on Global-Scale, the baseline dataset, and WorldRoadSeg-360K respectively, models evaluated on the same LSRV test set after 5 interaction steps obtain the following scores (Peng et al., 26 May 2026):

Pretrained dataset	Dice↑	clDice↑	APLS↑
Global-Scale	0.686	0.783	0.512
Baseline dataset	0.807	0.869	0.593
WorldRoadSeg-360K	0.835	0.905	0.620

The corresponding $0.8\text{–}1.1$9 and $512 \times 512$0 values are also lowest for WorldRoadSeg-360K pretraining, indicating fewer fragmented components and spurious loops (Peng et al., 26 May 2026). The paper interprets this as evidence that the diversity and mask-based supervision of WorldRoadSeg-360K provide a stronger pretraining signal for generalizable road extraction than large line-based datasets such as Global-Scale (Peng et al., 26 May 2026).

The benchmark is also evaluated in a practical annotation setting. In a user study on 100 images from WorldRoadSeg-360K, manual annotation achieved Dice 0.827, whereas RoadGIE-assisted annotation reached Dice 0.885 with about 7 interactions and 15 s per image, compared with 73 s for manual annotation (Peng et al., 26 May 2026). This corresponds to an approximately 79% reduction in annotation time (Peng et al., 26 May 2026). A plausible implication is that the dataset is intended not only as a passive benchmark but also as infrastructure for iterative dataset improvement.

The limitations are explicit. The benchmark targets imagery at $512 \times 512$1 m resolution, which may constrain transfer to much higher-resolution data or to other modalities such as SAR or multispectral imagery (Peng et al., 26 May 2026). Training is restricted by GPU memory to six interaction steps, with 5 steps used in experiments, so more complex inference-time interaction patterns may exceed the model’s typical training regime (Peng et al., 26 May 2026). The semi-automatic annotation process implies some level of label noise and dependence on OSM quality, yet the exact proportions of high- and low-quality subsets are not reported (Peng et al., 26 May 2026). The paper also does not quantify internal train/validation/test splits within WorldRoadSeg-360K, instead using it primarily as a large-scale training corpus with LSRV as an out-of-domain test set (Peng et al., 26 May 2026).

These omissions are best understood as under-specification rather than contradiction. WorldRoadSeg-360K is highly explicit about benchmark scale, topology-aware evaluation, and interactive usage, but less explicit about some conventional dataset-release details such as full metadata schema, licensing terms, and exhaustive split definitions (Peng et al., 26 May 2026). For researchers, that means its principal contribution lies in establishing a global-scale, mask-based benchmark for structurally faithful road extraction, while some operational details remain to be verified in the accompanying repository and supplementary materials (Peng et al., 26 May 2026).