Experimenting with lane segmentation. The first implementation is with vanilla UNET on the following dataset https://www.kaggle.com/thomasfermi/lane-detection-for-carla-driving-simulator.
The dataset is simple dataset from carla simulator.
Parameters that can be sent during training:
Learning Rate, --lr
Number of epochs, --epochs
Batch size, --batch_size
Resume training, --resume
Number of workers, --num_workers
Dataset path, --dataset_path
Example:
python train.py --dataset_path data/ --lr 0.001 --batch_size 16
python eval.py --image_path /path/to/image.png
- Add accuracy into training
- Improve tensorboard output: include more information, add images
- Add more information for saving checkpoints
- Add evaluation for the whole folder instead of image per image
- Extrapolate lines
- Mark driving lane
- Play around with augmentations, and generate bigger dataset
- Explore better solutions for this problem, like LaneNet
- Explore other datasets