BDD Lane Detection

Overview

A ResNet34-backbone U-Net model for detecting lanes using ~3k images from the BDD100K dataset. Built with modern PyTorch practices for maintainability and performance.

Features

Configuration Management

YAML-based configuration for easy parameter tuning
Organized into data, model, training, and inference sections

Pipeline Architecture

End-to-end training, evaluation, and inference
Modular components with automatic data splitting

Training Improvements

Checkpointing with optimizer state
Early stopping and LR scheduling

Visualization Tools

Training history and prediction overlays
Gradio web interface for demos

Project Structure

├── config.yaml Configuration ├── main.py CLI interface ├── app.py Gradio web app ├── data/ │ ├── images/ Training & validation images │ ├── labels/ Segmentation masks │ └── model.pt Pretrained checkpoint ├── src/ │ ├── model.py U-Net architecture │ ├── data.py Dataset classes │ ├── train.py Training loop │ ├── loss.py Dice+BCE loss │ └── pipeline.py End-to-end pipeline ├── source/ Notebook compatibility └── notebooks/ Jupyter experiments

Data and pretrained models available on Kaggle.

Notebooks

Notebook	Description
01-data	Dataset exploration and preprocessing
02-transform	Augmentation experiments
03-model	Model training
04-evaluate	Performance evaluation

Results

U-Net with ResNet-34 backbone trained using Dice + BCE loss, evaluated with IoU metric.

Training metrics showing loss and IoU curves

Full analysis in the PDF report.

Usage

Training

python main.py --mode train --config config.yaml

Evaluation

python main.py --mode evaluate --checkpoint checkpoints/model.pt

Prediction

python main.py --mode predict --input image.jpg --output predictions/

Web Demo

python app.py

Also hosted on Hugging Face Spaces.

Requirements

torch ≥2.0 torchvision numpy pandas opencv albumentations smp gradio ≥4.0 matplotlib pyyaml

pip install -r requirements.txt