Adapted from https://github.com/MaybeShewill-CV/lanenet-lane-detection and https://github.com/leonfrank/lanenet-danet-pytorch
Inspiration drawn from https://github.com/davidtvs/PyTorch-ENet https://github.com/sacmehta/ESPNet
Using ESPNet as Encoder-Decoder instead of ENet.
python setup.py install
To train on the test data included in the repo,
python3 lanenet/train.py --dataset ./data/training_data_example
Download TUsimple dataset from TuSimple/tusimple-benchmark#3
When done run the script in the scripts-folder (From https://github.com/MaybeShewill-CV/lanenet-lane-detection)
python tusimple_transform.py --src_dir <directory of downloaded tusimple>
After this run training as before:
python3 lanenet/train.py --dataset <tusimple_transform script output folder>
To train on a custom dataset, the easiest approach is to make sure it follows the format laid out in the data folder. Alternatively write a custom PyTorch dataset class (if you do, feel free to provide a PR)
Towards End-to-End Lane Detection: an Instance Segmentation Approach
https://arxiv.org/pdf/1802.05591.pdf
ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://arxiv.org/abs/1803.06815
ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation