Paper link

Video results

1. Installation
2. TuSimple
3. CULane
4. Unsupervised Llamas
5. Pre-trained Weights
6. Citation

1. Clone this repository
2. Install Anaconda
3. Create a virtual environment and install all dependencies:

conda create -n laneaf pip python=3.6
source activate laneaf
pip install numpy scipy matplotlib pillow scikit-learn
pip install opencv-python
pip install https://download.pytorch.org/whl/cu101/torch-1.7.0%2Bcu101-cp36-cp36m-linux_x86_64.whl
pip install https://download.pytorch.org/whl/cu101/torchvision-0.8.1%2Bcu101-cp36-cp36m-linux_x86_64.whl
source deactivate

You can alternately find your desired torch/torchvision wheel from here.

4. Clone and make DCNv2:

cd models/dla
git clone https://github.com/lbin/DCNv2.git
cd DCNv2
./make.sh

The entire TuSimple dataset should be downloaded and organized as follows:

└── TuSimple/
    ├── clips/
    |   └── .
    |   └── .
    ├── label_data_0313.json
    ├── label_data_0531.json
    ├── label_data_0601.json
    ├── test_tasks_0627.json
    ├── test_baseline.json
    └── test_label.json

The model requires ground truth segmentation labels during training. You can generate these for the entire dataset as follows:

source activate laneaf # activate virtual environment
python datasets/tusimple.py --dataset-dir=/path/to/TuSimple/
source deactivate # exit virtual environment

LaneAF models can be trained on the TuSimple dataset as follows:

source activate laneaf # activate virtual environment
python train_tusimple.py --dataset-dir=/path/to/TuSimple/ --backbone=dla34 --random-transforms
source deactivate # exit virtual environment

Other supported backbones are erfnet and enet.

Config files, logs, results and snapshots from running the above scripts will be stored in the LaneAF/experiments/tusimple folder by default.

Trained LaneAF models can be run on the TuSimple test set as follows:

source activate laneaf # activate virtual environment
python infer_tusimple.py --dataset-dir=/path/to/TuSimple/ --snapshot=/path/to/trained/model/snapshot --save-viz
source deactivate # exit virtual environment

This will generate outputs in the TuSimple format and also produce benchmark metrics using their official implementation.

The entire CULane dataset should be downloaded and organized as follows:

└── CULane/
    ├── driver_*_*frame/
    ├── laneseg_label_w16/
    ├── laneseg_label_w16_test/
    └── list/

LaneAF models can be trained on the CULane dataset as follows:

source activate laneaf # activate virtual environment
python train_culane.py --dataset-dir=/path/to/CULane/ --backbone=dla34 --random-transforms
source deactivate # exit virtual environment

Other supported backbones are erfnet and enet.

Config files, logs, results and snapshots from running the above scripts will be stored in the LaneAF/experiments/culane folder by default.

Trained LaneAF models can be run on the CULane test set as follows:

source activate laneaf # activate virtual environment
python infer_culane.py --dataset-dir=/path/to/CULane/ --snapshot=/path/to/trained/model/snapshot --save-viz
source deactivate # exit virtual environment

This will generate outputs in the CULane format. You can then use their official code to evaluate the model on the CULane benchmark.

The Unsupervised Llamas dataset should be downloaded and organized as follows:

└── Llamas/
    ├── color_images/
    |   ├── train/
    |   ├── valid/
    |   └── test/
    └── labels/
        ├── train/
        └── valid/

LaneAF models can be trained on the Llamas dataset as follows:

source activate laneaf # activate virtual environment
python train_llamas.py --dataset-dir=/path/to/Llamas/ --backbone=dla34 --random-transforms
source deactivate # exit virtual environment

Other supported backbones are erfnet and enet.

Config files, logs, results and snapshots from running the above scripts will be stored in the LaneAF/experiments/llamas folder by default.

Trained LaneAF models can be run on the Llamas test set as follows:

source activate laneaf # activate virtual environment
python infer_llamas.py --dataset-dir=/path/to/Llamas/ --snapshot=/path/to/trained/model/snapshot --save-viz
source deactivate # exit virtual environment

This will generate outputs in the CULane format and Llamas format for the Lane Approximations benchmark. Note that the results produced in the Llamas format could be inaccurate because we guess the IDs of the indivudal lanes.

You can download our pre-trained model weights using this link.

If you find our code and/or models useful in your research, please consider citing the following papers:

@article{abualsaud2021laneaf,
title={LaneAF: Robust Multi-Lane Detection with Affinity Fields},
author={Abualsaud, Hala and Liu, Sean and Lu, David and Situ, Kenny and Rangesh, Akshay and Trivedi, Mohan M},
journal={arXiv preprint arXiv:2103.12040},
year={2021}
}