docker

RoadTagger Docker Container

We containerize the pre_alpha_clean_version of RoadTagger into a docker container image. You can use this docker container to reproduce the numbers in the paper and run RoadTagger on custom images (please check out the usage below).

Usage

Docker

To install docker (with GPU support), you can check out the instruction here.

Inference on Custom Image

After the docker environment is set up, start the RoadTagger inference server.

docker run -p8010:8000 -p8011:8001 -it --rm songtaohe/roadtagger_inference_server_cpu:latest

To run RoadTagger on custom images, all you need is to provide an image and a graph file (edges in image coordinate), and specify the ground sampling distance (spatial resolution) of your image.

We show an example (image and graph file) in the examples folder. To run RoadTagger on this example,

cd script
python infer.py -image ../examples/e1.png -graph ../examples/e1.json -gsd 0.125

By default, this script will output the result to out.json. If everything goes well, the result should appear in out.json.

[
    {
    "node id": 0, 
    "road type": "primary", 
    "lane count probability": [
      0.01214602217078209, 
      0.09006574004888535, 
      0.810308039188385, 
      0.07358891516923904, 
      0.004188787192106247, 
      0.009702381677925587
    ], 
    "coordinate": [
      194, 
      176
    ], 
    "lane count": 3, 
    "road type probability": [
      0.34992465376853943, 
      0.6500753164291382
    ]
  }
]

Build Docker Container

If you want to build this docker container image by yourself, you have to first download the RoadTagger model so that it can work with custom images.

cd app/roadtagger
./downloadModel.sh

We have scripts to build and run the docker container on CPUs. You can also change the TensorFlow version in the Dockerfile to enable GPU support (the code should work with TensorFlow 1.15).