/cresi

Road network extraction from satellite imagery, with speed and travel time estimates

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

Alt text

CRESI

City-scale Road Extraction from Satellite Imagery

This repository provides an end-to-end pipeline to train models to detect routable road networks over entire cities, and also provide speed limits and travel time estimates for each roadway. We have observed success with both SpaceNet imagery and labels, as well as Google satellite imagery with OSM labels. The repository consists of pre-processing modules, deep learning segmentation model (inspired by the winning SpaceNet 3 submission by albu), post-proccessing modules to extract the road networks, inferred speed limits, and travel times. Furthermore, we include modules to scale up network detection to the city-scale, rather than just looking at small image chips. The output of CRESI is a geo-referenced NetworkX graph, with full access to the many graph-theoretic algorithms included in this package.
For further details see:

  1. Our WACV Paper
  2. Blogs:
    1. Large Road Networks
    2. Road Speeds
    3. OSM+Google Imagery
    4. SpaceNet 5 Baseline Part 1 - Data Prep
    5. SpaceNet 5 Baseline Part 2 - Segmentation
    6. SpaceNet 5 Baseline Part 3 - Road Graph + Speed
    7. SpaceNet 5 Speed / Performance Comparision

Install

  1. Download this repository

  2. Build docker image

     nvidia-docker build -t cresi /path_to_cresi/docker/[cpu, gpu]

  3. Create docker container (all commands should be run in this container)

     nvidia-docker run -it --rm -ti --ipc=host --name cresi_image cresi

Prep

  1. Prepare train/test data, e.g.:

     python /path_to_cresi/cresi/data_prep/speed_masks.py

  2. Edit .json file to select desired variables and point to appropriate directories

Train

  1. All at once

     cd /path_to_cresi/cresi

 ./train.sh configs/sn5_baseline.json

  2. Run commands individually

    A. Generate folds (within docker image)

     python /path_to_cresi/cresi/00_gen_folds.py configs/sn5_baseline.json

    B. Run train script (within docker image)

     python /path_to_cresi/cresi/01_train.py configs/sn5_baseline.json --fold=0

Test

  1. All at once

     cd /path_to_cresi/cresi

 ./test.sh configs/sn5_baseline.json

  2. Run commands individually

    A. Execute inference (within docker image)

     python /path_to_cresi/cresi/02_eval.py configs/sn5_baseline.json

    B. Merge predictions (if required)

     python /path_to_cresi/cresi/03a_merge_preds.py configs/sn5_baseline.json

    C. Stitch together mask windows (if required)

     python /path_to_cresi/cresi/03b_stitch.py configs/sn5_baseline.json

    D. Extract mask skeletons

     python /path_to_cresi/cresi/04_skeletonize.py configs/sn5_baseline.json

    E. Create graph

     python /path_to_cresi/cresi/05_wkt_to_G.py configs/sn5_baseline.json

    F. Infer road travel time and speed limit

     python /path_to_cresi/cresi/06_infer_speed.py configs/sn5_baseline.json

Outputs will look something like the image below:

Alt text