CARLA dataset generator

CARLA dataset generator uses CARLA simulator to generate videos of a car driving around the simulation environment. Four visual modalities (RGB, optical flow, depth, instance segmentation) are extracted, as well as the motion data of all vehicles and pedestrians at every timestep

The dataset generator can be parameterized with a CSV file (see generate_params.py) that contains a seed, the split name (train, test, validation), map name, framerate, duration, number of cars, number of pedestrians, weather and daytime, a speed limit factor, the cameras’ position, orientation, field of view, and resolution. The generator was designed to be fully deterministic.

One sample of the training set can be found at: https://tubcloud.tu-berlin.de/s/5ygs7dbMjDJWtqX

Note Please refer to my bachelor's thesis for details: Evaluating multi-stream networks for self-supervised representation learning

How to generate a CARLA dataset

1. Clone this repo
2. Go to the cloned folder

   cd <PATH_TO_THIS_REPO>

3. Get docker image with one of the following methods:

   #Pull image from dockerhub
   docker pull nschroeder/carla:0.9.13-v1
   
   #Compile image yourself
   wget https://carla-releases.s3.eu-west-3.amazonaws.com/Linux/AdditionalMaps_0.9.13.tar.gz
   docker build -t nschroeder/carla:0.9.13-v1 .

4. Edit src/generate_params.py which will later generate a .csv file, which defines all parameters for each sample. The current configuration creates the default4.csv which was used for the main dataset of this thesis.
5. Generate this .csv

   `docker run --rm -v <ABSOLUTE_PATH_TO_THIS_REPO>/src/:/mnt/scripts nschroeder/carla:0.9.13-v1 python3 generate_params.py <DATASET_NAME>`

6. Now a <DATASET_NAME>.csv should be in the <ABSOLUTE_PATH_TO_THIS_REPO>
7. Start the CARLA server with display or headless:

   # stop already running server containers  
   docker container stop csniklas && docker rm csniklas
       
   # with display
   docker run -it --privileged --gpus all --net=host --name=csniklas -e DISPLAY=$DISPLAY nschroeder/carla:0.9.13-v1 /home/carla/CarlaUE4.sh -nosound
       
   # without display
   docker run -it --restart=always --privileged --gpus=all --net=host --name=csniklas nschroeder/carla:0.9.13-v1 /home/carla/CarlaUE4.sh -RenderOffScreen -nosound

8. Open another terminal (also inside this repo folder)
9. Start the CARLA client:

   # stop already running client containers
   docker container stop ccniklas && docker rm ccniklas
   
   CODE_DIR=<ABSOLUTE_PATH_TO_THIS_REPO>/src/
   DATASET=<ABSOLUTE_PATH_TO_FOLDER_WHERE_DATASET_SHOULD_BE_SAVED>
   PARAM_NAME=<DATASET_NAME>
   
   docker run -it --net=host -v ${CODE_DIR}:/mnt/scripts -v ${DATASET}:/mnt/dataset -e PARAM_NAME=${PARAM_NAME} --name=ccniklas nschroeder/carla:0.9.13-v1 /bin/bash

10. Start datagenerator inside Dockercontainer

    ./auto_restart.sh client.py -f ${PARAM_NAME}.csv

11. The dataset will be written to the specified folder
12. Stop client and server

    docker container stop csniklas
    docker rm csniklas
    docker container stop ccniklas
    docker rm ccniklas

Entrypoint to code

The src/client.py first gets called.