RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction (CVPR 2026)

Chenxu Peng¹  Chenxu Wang¹  Yimian Dai^1,2  Yongxiang Liu³  Ming-Ming Cheng^1,2  Xiang Li^1,2* 

$^{1}$ VCIP, CS, Nankai University, $^{2}$ NKIARI, Shenzhen Futian, $^{3}$ National University of Defense Technology

📰Abstract

Accurate road segmentation from aerial imagery is fundamental to many geospatial applications. However, existing road segmentation datasets frequently exhibit imbalanced scene diversity, insufficient semantic granularity, and inadequate structural continuity, restricting their generalization capabilities across varied 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 provides a comprehensive benchmark for evaluating road segmentation models and highlights the challenges that current methods face 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 3.7 million parameters and real-time processing capabilities.

📰 News

• [2026.06.15] Released WorldRoadSeg-360K dataset V1.0.
• [2026.03.05] Released SOTA segmentation model weights pretrained on WorldRoadSeg-360K.
• [2026.02.25] Released model code and weights.
• [2026.02.21] Released demo.
• [2026.02.21] RoadGIE has been accepted to CVPR 2026! .

🚩Overview

RoadGIE supports connectivity-aware prompts, including clicks and scribbles, which inherently align with the topology of road networks.

Installation

# 1. download your environment
https://pan.baidu.com/s/1NblpjwPvn4PwVvhqfvwT7A code:8sj2 
# 2. make a directory and unzip your environment
cd RoadGIE
mkdir -p env
tar -xzf RoadGIE.tar.gz -C env
# 3. activate environment
source env/bin/activate

Inference

# 1. run the demo
python demo/demo.py
# 2.open the gradio demo
http://192.168.4.12:7860/

Data

Please download the dataset via Baidu Cloud(ebh5).

Training

python RoadGIE/roadgie/experiment/unet.py -config train_unet.yaml 

SOTA segmentation model weights pretrained on WorldRoadSeg-360K

Encoder	Decoder	Dice	Model	Weight
SegNeXt_base	DaFormer33_ASPP	0.740	SegNeXt_base	SegNeXt_base
tf_efficientnetv2_l_in21ft1k	SiUnet	0.726	tf_efficientnetv2_l	tf_efficientnetv2_l
tf_efficientnet_b7_ns	SiUnet	0.719	tf_efficientnet_b7_ns	tf_efficientnet_b7_ns
convnext_large_384_in22ft1k	SiUnet	0.740	convnext_large_384	convnext_large_384
SwinT_small	UperNet	0.737	SwinT_small	SwinT_small
PvT_v2_b4	DaFormer33_ASPP	0.738	PvT_v2_b4	PvT_v2_b4
Dual_ViT_b	DaFormer33_ASPP	0.742	Dual_ViT_b	Dual_ViT_b
SMT_base	DaFormer33_ASPP	0.742	SMT_base	SMT_base
Uniformer_base	DaFormer33_ASPP	0.740	Uniformer_base	Uniformer_base
WaveViT	DaFormer33_ASPP	0.738	WaveViT	WaveViT
MiT_b2	SegFormer	0.735	MiT_b2	MiT_b2
iFormer_large	DaFormer33_ASPP	0.739	iFormer_large	iFormer_large
HorNet_small_gf	DaFormer33_ASPP	0.742	HorNet_small_gf	HorNet_small_gf
CoaT_4level_small	DaFormer33_ASPP	0.736	CoaT_4level_small	CoaT_4level_small
CoaT_4level_lite_medium	DaFormer33_ASPP	0.739	CoaT_4level_lite	CoaT_4level_lite
CoaT_5level_parallel_small	DaFormer33_ASPP	0.741	CoaT_5level_parallel	CoaT_5level_parallel

Results

Comparison of different models using different datasets.

Method	Baseline dataset (Dice↑)	Baseline dataset (APLS↑)	WorldRoadSeg-360K (Dice↑)	WorldRoadSeg-360K (APLS↑)
EISeg	0.701	0.511	0.706	0.515
ScribbleSeg-B0	0.766	0.560	0.785	0.578
ScribbleSeg-B3	0.761	0.556	0.788	0.580
SAM (ViT-b)	0.719	0.522	0.737	0.539
SAM (ViT-h)	0.738	0.539	0.756	0.553
PRISM-2D	0.622	0.463	0.643	0.481
PRISM-2D-Lite	0.656	0.489	0.669	0.496
ScribblePrompt	0.791	0.584	0.809	0.592
RoadGIE	0.807	0.593	0.835	0.620

Performance of models pretrained on different datasets and evaluated on the same test set. Best results are highlighted in bold.

Pretrained dataset	Dice↑	Recall↑	clDice↑	APLS↑	β₀↓	β₁↓
Baseline dataset	0.807	0.897	0.869	0.593	8.150	3.061
WorldRoadSeg-360K	0.835	0.934	0.905	0.620	5.823	2.752

To Do

• Release demo
• Release model code
• Release model weights
• Release training code
• Release WorldRoadSeg-360K dataset

Acknowledgements

• Our training code builds on the ScribblePrompt library. Thanks to @halleewong for sharing this code!

Citation

If you use this in your research, please cite this project.

@inproceedings{peng2026roadgie,
  title={RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction},
  author={Peng, Chenxu and Wang, Chenxu and Dai, Yimian and Liu, Yongxiang and Cheng, Ming-Ming and Li, Xiang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13285--13295},
  year={2026}
}