Roadmap Generation Using a Multi-Stage Ensemble of Deep Neural Networks With Smoothing-Based Optimization

Road detection from aerial images is a challenging task for humans and machines alike. Occlusion, the lack of visual cues and slim class borders for other road-like structures (such as pathways or private alleys) make the problem inherently ambiguous, requiring logic that goes beyond the input image. We propose a three-stage method for the task of road segmentation - first, an ensemble of multiple U-Net like CNNs generate binary road masks. Second, another CNN learns to refine roads segmentations based on the fusion of the road maps from the first stage. Third, missing links are added based on the inferred graph to improve segmentation.

@inproceedings{costea2018roadmap,
  title = {Roadmap Generation Using a Multi-Stage Ensemble of Deep Neural Networks With Smoothing-Based Optimization},
  author = {Dragos Costea and Alina Marcu and Emil Slusanschi and Marius Leordeanu},
  year = {2018},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
}