Roadline Semantic Segmentation

Semantic segmentation model based on U-Net with ResNet50 encoder.

🗃️ Training Data

Training data consists of 202 images which are frames from dash cameras videos. Labeling was performed using the Supervisely service.

❓ How to use

This guide is for Linux, however project works in Windows 10 and higher as well, just make sure you have PyTorch and CUDA installed in your system. Python 3.11.x is preferred.

1. Clone repository

git clone https://github.com/dead4an/roadline-semantic-segmentation.git
cd roadline-semantic-segmentation

2. Install requirements

Linux

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Windows

python3 -m venv .venv
./.venv/scripts/activate
pip install -r requirements.txt

3. Run train_model.py script

python3 train_model.py

In case of troubles due to CUDA out of memory I recommend to change BATCH_SIZE parameter in train_model.py to lower values, also you can lower resize_shape parameter in train_dataset initialization.

4. Inference

You can find example of inference in inference.ipynb.

💡 Improvement ideas

• Extend training dataset
• Add augmentation
• Test other encoders