A baseline solution to the 2018 VIAME detection challenge

This repo outlines a baseline solution to the 2018 VIAME Detection Challenge using the algorithms provided by Detectron system (developed by Facebook Research).

Challenge Website: 1

The instructions in this script rely on a few predefined directories. You may overwrite these to fit your personal workflow.

CODE_DIR=$HOME/code
DATA_DIR=$HOME/data
WORK_DIR=$HOME/work

First, download the groundtruth (phase0-annotations.tar.gz) and the images (phase0-imagery.tar.gz) from 2.

After downloading the data from challenge.kitware.com, extract it to your data directory

mkdir -p $DATA_DIR/viame-challenge-2018
tar xvzf $HOME/Downloads/phase0-annotations.tar.gz -C $DATA_DIR/viame-challenge-2018
tar xvzf $HOME/Downloads/phase0-imagery.tar.gz -C $DATA_DIR/viame-challenge-2018

tar xvzf data-challenge-training-imagery.tar.gz
tar xvzf test_data.tar.gz

Assuming you already have installed nvidia-docker, clone the Detectron repo and build the associated docker image.

DETECTRON=$CODE_DIR/Detectron
if [ ! -d "$DETECTRON" ]; then
    git clone https://github.com/facebookresearch/Detectron.git $DETECTRON
fi
# Build the docker container with caffe2 and detectron (which must use python2 ☹)
cd $DETECTRON/docker
docker build -t detectron:c2-cuda9-cudnn7 .
# test the image to make sure it works
nvidia-docker run -v ~/data:/data --rm -it detectron:c2-cuda9-cudnn7 python2 tests/test_batch_permutation_op.py