To train the model on the CULane dataset, ensure the frame difference .npz file is located at dataset/culane/list/train_diffs.npz. Then, run the training script using the desired configuration file.

python tools/train.py configs/clrernet/culane/clrernet_culane_dla34.py

Use docker-compose to build the development environment with your current user's UID and then run the container. This ensures file permissions are handled correctly within the container.

docker-compose build --build-arg UID="`id -u`" dev
docker-compose run --rm dev

To avoid overfitting during training, CLRerNet uses a frame difference calculation to filter out redundant frames. This script outputs an .npz file containing the difference values.

python tools/calculate_frame_diff.py [culane_root_path]

Before running inference or testing, download the pretrained weights for the CULane benchmark. You can choose between the standard CLRerNet or the version with Exponential Moving Average (EMA).

CLRerNet:

wget https://github.com/hirotomusiker/CLRerNet/releases/download/v0.1.0/clrernet_culane_dla34.pth

CLRerNet\u2082 (w/ EMA):

wget https://github.com/hirotomusiker/CLRerNet/releases/download/v0.1.0/clrernet_culane_dla34_ema.pth

If you do not have a docker-compose environment, install it and start the docker service using apt.

To install the NMS (Non-Maximum Suppression) extension, you must ensure the nvidia runtime is enabled during the build process. This is done by configuring /etc/docker/daemon.json to set nvidia as the default-runtime.

# Install docker-compose and start service
sudo apt install docker-compose
sudo service docker start

Configure /etc/docker/daemon.json

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

Restart docker to apply changes

sudo systemctl restart docker

The recommended way to install CLRerNet is via a Docker environment. This repository is based on the mmdetection 3.3 environment. If you require the older mmdet2x-based version, you must checkout the v0.2.1 branch.

To build and run the development environment, use the following commands:

docker compose build --build-arg UID="`id -u`" dev
docker compose run --rm dev

If the NMS extension installation fails during the Docker build process, you can manually install it inside the running container.

1. Modify the Dockerfile to skip the line: RUN python /tmp/nms/setup.py install.
2. Build and run the container.
3. Once inside the container, navigate to the NMS directory and run the installation manually.

# Inside the container
cd libs/models/layers/nms/
python setup.py install
cd /work

To evaluate the model's performance on the CULane dataset, run the tools/test.py script with the appropriate config and checkpoint files. Ensure you have followed the Dataset Preparation guide first.

python tools/test.py configs/clrernet/culane/clrernet_culane_dla34_ema.py clrernet_culane_dla34_ema.pth

To detect lanes in a single image and visualize the results, use the demo/image_demo.py script. You must provide the configuration file, the pretrained checkpoint, and specify an output file path.

python demo/image_demo.py demo/demo.jpg configs/clrernet/culane/clrernet_culane_dla34_ema.py clrernet_culane_dla34_ema.pth --out-file=result.png

You can calculate the frames per second (FPS) by measuring inference iterations. Use the tools/speed_test.py script, specifying the config, checkpoint, a sample image, and the number of warmup and test iterations.

python tools/speed_test.py configs/clrernet/culane/clrernet_culane_dla34.py clrernet_culane_dla34.pth --filename demo/demo.jpg --n_iter_warmup 1000 --n_iter_test 10000