SPIN Road Mapper: Extracting Roads from Aerial Images via Spatial and Interaction Space Graph Reasoning for Autonomous Driving

Wele Gedara Chaminda Bandara, Jeya Maria Jose Valanarasu, Vishal M. Patel

Abstract

Road extraction is an essential step in building autonomous navigation systems. Detecting road segments is challenging as they are of varying widths, bifurcated throughout the image, and are often occluded by terrain, cloud, or other weather conditions. Using just convolution neural networks (ConvNets) for this problem is not effective as it is inefficient at capturing distant dependencies between road segments in the image which is essential to extract road connectivity. To this end, we propose a Spatial and Interaction Space Graph Reasoning (SPIN) module which when plugged into a ConvNet performs reasoning over graphs constructed on spatial and interaction spaces projected from the feature maps. Reasoning over spatial space extracts dependencies between different spatial regions and other contextual information. Reasoning over a projected interaction space helps in appropriate delineation of roads from other topographies present in the image. Thus, SPIN extracts long-range dependencies between road segments and effectively delineates roads from other semantics. We also introduce a SPIN pyramid which performs SPIN graph reasoning across multiple scales to extract multi-scale features. We propose a network based on stacked hourglass modules and SPIN pyramid for road segmentation which achieves better performance compared to existing methods. Moreover, our method is computationally efficient and significantly boosts the convergence speed during training, making it feasible for applying on large-scale high-resolution aerial images.

The Challenge of Road Extraction

Accurate road maps are foundational for autonomous driving, disaster response, and automated navigation. However, extracting continuous road networks from high-resolution aerial imagery is a complex semantic segmentation task. Roads vary significantly in width, branch out across vast areas, and are frequently obscured by tree canopies, building shadows, or clouds.

Standard Convolutional Neural Networks (ConvNets) are limited by their localized receptive fields. This inherent "inductive bias" makes them struggle with long-range dependencies—the ability to recognize that two visible road segments separated by an occlusion are part of the same continuous path. While Vision Transformers (ViTs) can model these dependencies, their computational cost is often prohibitive for the massive scale of high-resolution satellite data.

Methodology: The SPIN Module

The core innovation of this work is the SPIN (Spatial and Interaction Space Graph Reasoning) module. Designed as a lightweight plug-in for ConvNets, it provides global context by performing graph reasoning in two distinct domains:

  1. Spatial Space Graph Reasoning: The module constructs a fully connected graph over the spatial feature map. By computing similarities between distant pixels, it explicitly models long-range connectivity, allowing the network to "bridge the gap" between disconnected road segments.
  2. Interaction Space Graph Reasoning: Features are projected into a latent "interaction space" where pixels with similar semantics (e.g., all road-like features vs. all vegetation) are grouped into nodes. Reasoning in this space helps the model distinguish roads from topographically similar features, such as building roofs.
Conceptual overview of the SPIN module, illustrating how spatial and interaction space reasoning helps delineate road segments.

To handle roads of varying widths, the authors implement a SPIN Pyramid, which applies this dual-graph reasoning across multiple feature scales (1x, 0.5x, and 0.25x). The final architecture utilizes a stacked hourglass backbone and a multi-task learning approach to predict both road segmentation and road orientation simultaneously.

Key Insights and Efficiency

A significant advantage of the SPIN module is its efficiency. Despite providing global context similar to a Transformer, it adds only 0.03M parameters to the base network. Furthermore, the inclusion of global context significantly accelerates training; the authors observed that the SPIN-enhanced model converges nearly twice as fast as the baseline ConvNet. This efficiency makes it highly suitable for processing large-scale, high-resolution remote sensing datasets.

Results and Performance

The SPIN Road Mapper was evaluated on the Massachusetts Roads and DeepGlobe datasets, achieving state-of-the-art (SOTA) results on both.

  • Massachusetts Roads: Achieved an F1 score of 84.47% and an APLS (Average Path Length Similarity) of 72.49%. The APLS metric is particularly important as it specifically measures the topological connectivity of the extracted road graph.
  • DeepGlobe: Reached an F1 score of 84.32% and an APLS of 74.14%.

Qualitative analysis confirms that the SPIN module effectively maintains road connectivity in areas where baseline models like SegNet or LinkNet produce fragmented results due to shadows or tree cover.

Visual comparison of road extraction results. Red boxes highlight areas where SPIN Road Mapper successfully maintains connectivity despite occlusions, outperforming existing baselines.

Conclusion and Future Work

The SPIN Road Mapper demonstrates that graph reasoning is a powerful, lightweight alternative to Transformers for capturing long-range dependencies in spatial data. While the model currently performs best with relatively clean training data—the authors noted that removing heavily occluded images from the training set improved performance—its computational efficiency and superior connectivity metrics make it a promising candidate for real-world autonomous navigation and large-scale mapping applications.

Summary updated

Abstract

Road extraction is an essential step in building autonomous navigation systems. Detecting road segments is challenging as they are of varying widths, bifurcated throughout the image, and are often occluded by terrain, cloud, or other weather conditions. Using just convolution neural networks (ConvNets) for this problem is not effective as it is inefficient at capturing distant dependencies between road segments in the image which is essential to extract road connectivity. To this end, we propose a Spatial and Interaction Space Graph Reasoning (SPIN) module which when plugged into a ConvNet performs reasoning over graphs constructed on spatial and interaction spaces projected from the feature maps. Reasoning over spatial space extracts dependencies between different spatial regions and other contextual information. Reasoning over a projected interaction space helps in appropriate delineation of roads from other topographies present in the image. Thus, SPIN extracts long-range dependencies between road segments and effectively delineates roads from other semantics. We also introduce a SPIN pyramid which performs SPIN graph reasoning across multiple scales to extract multi-scale features. We propose a network based on stacked hourglass modules and SPIN pyramid for road segmentation which achieves better performance compared to existing methods. Moreover, our method is computationally efficient and significantly boosts the convergence speed during training, making it feasible for applying on large-scale high-resolution aerial images. Code available at: https://github.com/wgcban/SPIN_RoadMapper.git.
BibTeX
@inproceedings{bandara2022spin,
  author = {Wele Gedara Chaminda Bandara and Jeya Maria Jose Valanarasu and Vishal M. Patel},
  title = {{SPIN Road Mapper: Extracting Roads from Aerial Images via Spatial and Interaction Space Graph Reasoning for Autonomous Driving}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2022},
  url = {https://openreview.net/forum?id=EC8ZhYU_mf},
}
Figure