Cityscape Segmentation using Segformer Architecture

Project Overview

This project implements semantic segmentation on the Cityscape dataset using the Segformer architecture. The model is trained to segment urban scenes into various classes, including roads, buildings, cars, pedestrians, and 19 such classes in total.

Table of Contents

• Introduction
• Dataset
• Architecture

Introduction

Semantic segmentation is a pixel-level classification task aimed at assigning a class label to every pixel in an image. This project uses the Segformer architecture to perform segmentation on the Cityscape dataset, which contains images of urban environments.

Dataset

At the time of this project a total of 19 classes were considered, There were 2380 train images, 595 validation images, 500 test Images For more information on the dataset, visit the official Cityscape Dataset.

Architecture

Segformer is a transformer-based model designed for semantic segmentation tasks. It leverages both local and global feature representations, providing a robust and scalable solution for vision tasks. This project uses the Segformer implementation from scratch, the model was built using the pytorch modular approach.

For more details, refer to the original author of this architecture here.