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

TL;DR

Extracting continuous road networks from aerial imagery is notoriously difficult for standard ConvNets due to occlusions and the need to understand long-range dependencies. This paper introduces the SPIN (Spatial and Interaction Space Graph Reasoning) module, a lightweight plug-in that uses graph convolutions in both spatial and latent semantic spaces to better connect road segments and separate them from background objects. The resulting "SPIN Road Mapper" achieves state-of-the-art accuracy, improves road connectivity, and converges twice as fast as baseline models.

What Problem Does This Paper Address?

Autonomous driving, disaster mitigation, and automated navigation systems rely heavily on accurate road maps. Automatically extracting these roads from aerial images is treated as a semantic segmentation problem, but it comes with unique challenges: roads vary wildly in width, branch out continuously across the entire image, and frequently get occluded by clouds, tree canopies, or building shadows.

Standard Convolutional Neural Networks (ConvNets) struggle here because their localized receptive fields (inductive biases) make them poor at learning the long-range dependencies required to recognize that two disconnected visible road segments are actually part of the same continuous street. While Vision Transformers excel at long-range dependencies, they are far too computationally expensive to run on massive, high-resolution aerial images. The authors aim to bridge this gap by using highly efficient graph reasoning to give ConvNets a global understanding of the image.

Methods

The core contribution is the SPIN (Spatial and Interaction Space Graph Reasoning) module, which can be plugged into a standard ConvNet. It performs graph reasoning in two distinct domains:

  1. Spatial Space Graph Reasoning: The model builds a fully connected graph over the original spatial feature map. By calculating dot-product similarities between distant pixels, it explicitly models long-range connectivity, helping the network "connect the dots" between distant road segments.
  2. Interaction Space Graph Reasoning: The model projects the spatial features into a latent "interaction space." In this space, pixels with similar semantics (e.g., all road pixels, all tree pixels) are grouped together into nodes, regardless of where they are in the image. Graph reasoning here helps cleanly delineate road features from non-road topographies. The updated features are then projected back to the original spatial dimensions.

To capture roads of varying widths, the authors introduce a SPIN Pyramid, which applies this dual-graph reasoning across multiple feature scales (1, 1/2, and 1/4). The overall network architecture uses a stacked hourglass backbone and employs multi-task learning to predict both road segmentation and road orientation simultaneously.

Datasets

The model was trained and evaluated on two large-scale, standard benchmarks:

  • Massachusetts Roads Dataset: Contains 1500 × 1500 pixel images, cropped into 512 × 512 patches. Note: The authors manually removed heavily occluded images from the training set, as they found these severely degraded model performance.
  • DeepGlobe Dataset: Contains high-resolution 1024 × 1024 satellite images, also processed into 512 × 512 overlapping patches for training.

Key Insights and Ideas

  • Dual-Space Reasoning is Synergistic: Spatial reasoning ensures the road network is connected, while interaction reasoning ensures the road isn't confused with a similarly colored building roof. Combining them yields significantly better results than using either alone.
  • Graph Convolutions as a Lightweight Transformer Alternative: Graph reasoning provides the global context and long-range dependency modeling of Vision Transformers but at a fraction of the computational cost. The SPIN module adds a mere 0.03M parameters to the base network.
  • Context Speeds Up Convergence: Because the SPIN module explicitly feeds the network highly contextual, global information, the network figures out the road topology much faster. The authors observed that adding SPIN cuts the required training time (convergence rate) in half.

Key Results

The SPIN Road Mapper was compared against strong baselines including U-Net, LinkNet34, Stack-HourGlass, and Batra et al.

  • Quantitative: It achieved State-of-the-Art (SOTA) on both datasets.
  • Massachusetts: Reached an F1 score of 84.47% and an APLS (Average Path Length Similarity—a metric specifically measuring graph/road connectivity) of 72.49% (+1.15% over the previous SOTA).
  • DeepGlobe: Reached an F1 score of 84.32% and an APLS of 74.14% (+1.02% over the previous SOTA).
  • Qualitative: Visualizations show the SPIN Road Mapper successfully bridging gaps in roads caused by tree canopy or building shadows, whereas baseline models predict broken, disconnected lines.

Limitations and Future Work

  • Limitations: The authors do not include a dedicated limitations section. However, their methodology notes that they had to manually remove heavily occluded images from the Massachusetts training set. This implies the model still relies on relatively clean training data to learn optimal representations and may struggle to learn if the ground truth is too noisy or hidden.
  • Future Work: While not explicitly detailed, the authors emphasize that the SPIN module's computational efficiency and fast convergence make it highly promising for broader applications in training ConvNets on large-scale, high-resolution remote sensing datasets beyond just road extraction.

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},
}

Loading related content...

Figure