Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method

Pan Yin,Kaiyu Li,Xiangyong Cao,Jing Yao,Lei Liu,Xueru Bai,Feng Zhou,Deyu Meng

2024-11-23

Abstract:Recently, road graph extraction has garnered increasing attention due to its crucial role in autonomous driving, navigation, etc. However, accurately and efficiently extracting road graphs remains a persistent challenge, primarily due to the severe scarcity of labeled data. To address this limitation, we collect a global-scale satellite road graph extraction dataset, i.e. Global-Scale dataset. Specifically, the Global-Scale dataset is $\sim20 \times$ larger than the largest existing public road extraction dataset and spans over 13,800 $km^2$ globally. Additionally, we develop a novel road graph extraction model, i.e. SAM-Road++, which adopts a node-guided resampling method to alleviate the mismatch issue between training and inference in SAM-Road, a pioneering state-of-the-art road graph extraction model. Furthermore, we propose a simple yet effective ``extended-line'' strategy in SAM-Road++ to mitigate the occlusion issue on the road. Extensive experiments demonstrate the validity of the collected Global-Scale dataset and the proposed SAM-Road++ method, particularly highlighting its superior predictive power in unseen regions. The dataset and code are available at \url{<a class="link-external link-https" href="https://github.com/earth-insights/samroadplus" rel="external noopener nofollow">this https URL</a>}.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problems that this paper attempts to solve mainly focus on two aspects:

1. **Limitations of the dataset**: One of the main challenges faced by existing road map extraction methods is the lack of comprehensive datasets and benchmark tests. The existing graph - label datasets are small in scale and mainly concentrated in urban areas, lacking coverage of complex environments such as rural and mountainous areas. In addition, the existing datasets cannot meet the needs of modern models in terms of scale and diversity, which leads to inaccurate algorithm evaluation and challenges in model generalization ability.

2. **Mismatch problems in the model training and inference process**: Some existing road map extraction models (such as SAM - Road) rely on node labels during the training phase, but use nodes selected from the predicted road masks during the inference process. This mismatch leads to inconsistent behavior of the model in the training and inference stages, affecting the performance of the model.

To solve these problems, the paper proposes the following solutions:

- **Construct a global - scale dataset**: The paper has collected a new global - scale road map extraction dataset - Global - Scale. This dataset is 20 times larger than the existing public datasets, covering an area of more than 13,800 square kilometers around the world, including various complex environments such as urban, rural, and mountainous areas. The dataset is divided into training sets, validation sets, and test sets, and also includes an unseen - domain test set to evaluate the generalization ability of the model in different geographical environments.

- **Propose a new model SAM - Road++**: This model solves the mismatch problem in the training and inference process by introducing the node - guided resampling strategy. Specifically, during the training phase, the model resamples nodes from the predicted road masks, making the node selection in the training process closer to that in the inference process. In addition, the paper also proposes the "extended - line" strategy, which uses the extended - line information between nodes to alleviate the road occlusion problem and improve the performance of the model in complex environments.

Through these improvements, the paper not only provides a more comprehensive dataset but also proposes a more effective road map extraction model, significantly improving the performance and generalization ability of the road map extraction task.