watonomous / deft

Joint detection and tracking model named DEFT, or ``Detection Embeddings for Tracking." Our approach relies on an appearance-based object matching network jointly-learned with an underlying object detection network. An LSTM is also added to capture motion constraints.

DEFT - WATonomous

This repo is derived from the original DEFT repo and our CentreTrack fork

The original readme can be found in README.original.md

Getting Started

1. Initialize submodules:

git submodule update --init --recursive

2. Make sure /etc/docker/daemon.json contains "default-runtime": "nvidia" as a top-level property. For example:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

If changes are made, restart the docker daemon:

sudo service docker restart

3. Start the container. Some assumptions have been made in docker-compose.yml (e.g. where the data is stored). Edit the file if needed.

./generate-dot-env.sh
HOST=$(hostname) docker-compose up -d

4. Open a shell

docker-compose exec dev /bin/bash
cd /project

6. Run DEFT

cd src
python test.py tracking,ddd --load_model models/DEFT/model_kitti.pth --dataset kitti_tracking --save_video --test --test_dataset kitti_tracking