Embedded Files
Motivation
Motivation
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 an interaction space helps in the 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.
An overview of our proposed method. We build graphs in two spaces: (a) spatial space and (b) a projected latent interaction space from feature maps. Graph reasoning in spatial space extracts connectivity between the road segments, whereas reasoning over interaction space delineates roads from other topographies. Nodes connected with lines in (a) denote how road segments are modeled to understand connectivity in the spatial space. Regions marked with different colors in (b) denote how different semantics are segregated for better road delineation in the interaction space.
SPIN Architecture
SPIN Architecture
The architecture of our proposed method. (a) We perform graph reasoning in spatial and interaction space. (b) The proposed SPIN pyramid module which performs SPIN graph reasoning at multiple scales (1, 1/2, and 1/4) of original feature map to extract multi-scale long-range contextual information.
Road Segmentation Network
Road Segmentation Network
Proposed network for road segmentation from aerial images. The input images are first feed forwarded to a feature extractor block followed by a bottleneck consisting of stack of two hourglass modules. Then, the output of bottleneck is passed through a segmentation branch which consists of conv layers, our SPIN pyramid and a final classification layer to get the road segmentation map.
Results
Results
Quantitative Results
Quantitative Results
Qualitative Results
Qualitative Results
Ablation Study
Ablation Study
We conduct an ablation study to demonstrate the effect of spatial, interaction, and SPIN graph reasoning on road segmentation.
Citation
Citation
@article{bandara2021spin,
title={SPIN Road Mapper: Extracting Roads from Aerial Images via Spatial and Interaction Space Graph Reasoning for Autonomous Driving},
author={Bandara, Wele Gedara Chaminda and Valanarasu, Jeya Maria Jose and Patel, Vishal M},
journal={arXiv preprint arXiv:2109.07701},
year={2021}
}
Page updated
Google Sites
Report abuse