carlesventura/iterative-deep-learning

Iterative Deep Learning for Road Topology Extraction

Published at BMVC 2018.

Paper available on ArXiv: https://arxiv.org/abs/1808.09814

Paper website: https://carlesventura.github.io/iterative-dl-road-website/

Code instructions

Downloading datasets:

Download Massachusetts Roads Dataset from website: https://www.cs.toronto.edu/~vmnih/data/

Download DRIVE dataset (vessels from retina images) from website: https://www.isi.uu.nl/Research/Databases/DRIVE/

Download graph annotations for DRIVE dataset from website: http://people.duke.edu/~sf59/Estrada_TMI_2015_dataset.htm

Set your work directory, create a directory inside named gt_dbs and copy there the downloaded datasets (roads dataset in a folder named MassachusettsRoads, DRIVE dataset in a folder named DRIVE and graph annotations for DRIVE in a folder named artery-vein).

Experiments for road topology extraction:

1. Generate road patches for training the patch-level model: roads/patch/generate_gt_val_roads.py
2. Train patch-level model: roads/patch/train_road_patches.py
3. (Optional) Evaluate patch-level model: roads/patch/evaluation/PR_evaluation_patch_roads.py
4. Apply the patch-level model iteratively over the road test images: roads/iterative/iterative_roads_local_mask.py
5. (Optional) Evaluate iterative results: roads/iterative/evaluation/connectivity_evaluation_roads.py

Experiments for vessel topology extraction:

1. Train patch-level model: vessels/patch/train_hg.py
2. (Optional) Evaluate patch-level model: vessels/patch/evaluation/PR_evaluation.py
3. Apply the patch-level model iteratively over the retina test images: vessels/iterative/iterative_graph_creation_no_mask_offset.py
4. (Optional) Evaluate iterative results: vessels/iterative/evaluation/connectivity_evaluation.py