Roadmap Generation using a Multi-Stage Ensemble of DNNs with SBO

Team members

Alina Marcu, Dragos Costea

Supervisors

Prof. dr. Marius Leordeanu

Prof. dr. Emil Slusanschi


Idea

Use ensemble of various dilation rates networks for better pixelwise detection + SBO for graph creation

Method

We developed the following pipeline

  • Extract road segmentation maps from trained U-Net-like networks
  • Detect intersections
  • Concatenate our networks results (for both roads and intersections)
  • Train another network on top of previous predictions
  • Create road graph from the final predictions
  • Add missing links inferred from the graph

Stage 1: We detect roads using an ensemble of U-nets with variable dilation rates for the bottleneck:

Max dilation 32 (1, 2, 4, 8, 16, 32)

Max dilation 48 (1, 2, 4, 8, 16, 32, 48)

Max dilation 64 (1, 2, 4, 8, 16, 32, 48, 64)

Stage 2: The generated road maps, along with the RGB and detected intersections are concatenated and fed to the next stage U-net-like architecture trained for road segmentation.

Stage 3: We generate road vectors from the improved segmentation map and further use these vectors to add missing links.


A . RGB input

B. Ground truth

C. Ensemble 1 (Stage 1), shown as sum of 4 CNN outputs:

  • Max dilation 32 - blue
  • Max dilation 48 - green
  • Max dilation 64 - red
  • Max dilation 32, same width thin - grey

D. Ensemble 2, the output of Stage 2


Different dilation rates generate different road predictions. Thin roads hurt training:

More is not better if the detected roads are not labeled... See the ground truth in the image above.

Road graphs help add missing links:

A .RGB input

B . binary mask from CNN

C. plotted road vectors from SBO

D. final segmentation with missing links added to the road vectors

Final results on comeptition's testing set:

Cite

Costea, Dragos, et al. "Roadmap Generation using a Multi-Stage Ensemble of Deep Neural Networks with Smoothing-Based Optimization." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2018.

@inproceedings{costea2018roadmap,
  title={Roadmap Generation using a Multi-Stage Ensemble of Deep Neural Networks with Smoothing-Based Optimization},
  author={Costea, Dragos and Marcu, Alina and Slusanschi, Emil and Leordeanu, Marius},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  pages={220--224},
  year={2018}
}