LaneAF: Robust Multi-Lane Detection with Affinity Fields

Paper link

Video results

Overview

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

Installation

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

TuSimple

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

Training

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.

Inference

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.

Results

Backbone	F1-score	Accuracy	FP	FN
DLA-34	96.4891	95.6172	0.0280	0.0418
ERFNet	94.9465	95.2978	0.0550	0.0465
ENet	92.8905	94.7271	0.0885	0.0560

CULane

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

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

Training

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.

Inference

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.

Results

Backbone	Total	Normal	Crowded	Dazzle	Shadow	No line	Arrow	Curve	Cross	Night
DLA-34	77.41	91.80	75.61	71.78	79.12	51.38	86.88	71.70	1360	73.03
ERFNet	75.63	91.10	73.32	69.71	75.81	50.62	86.86	65.02	1844	70.90
ENet	74.24	90.12	72.19	68.70	76.34	49.13	85.13	64.40	1934	68.67

Unsupervised Llamas

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

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

Training

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.

Inference

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.

Results

Backbone	F1-score	Precision	Recall	TP	FP	FN
DLA-34	96.01	96.91	95.26	71793	2291	3576
ERFNet	NA	NA	NA	NA	NA	NA
ENet	NA	NA	NA	NA	NA	NA

Pre-trained Weights

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

Citation

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}
}