Road Segmentation using DeepLabV3+

Overview

This project implements Road Segmentation using Deep Learning with DeepLabV3+ (ResNet101 backbone) in PyTorch. The model performs pixel-wise classification to detect road areas from urban scene images. This project demonstrates:

Semantic Segmentation
Transfer Learning
Custom Dataset Handling
Dice + BCE Loss Combination
IoU Evaluation Metric
Visualization with Overlay

goal

The goal is to accurately segment road regions from street-level images for applications such as:

Autonomous Driving
Drone Navigation
Intelligent Transportation Systems
Real-time Road Understanding

Requirements

torch torchvision transforms opencv-python matplotlib

Model Architecture

DeepLabV3+
Backbone: ResNet101
Pretrained on ImageNet
Final classifier modified for binary segmentation (road vs background)

Key Features

Transfer Learning Custom Dataset Class Augmentation using Albumentations Pixel-wise Binary Segmentation Overlay Visualization Modular and Clean Code Structure

How to Train

python -u "segmentation_road.py"

How to Evaluate and Predict

python -u "eval_predict.py"