Deep Learning for Visual Computing (DLVC)

This repository contains implementations and experiments for a Deep Learning for Visual Computing course, covering image classification and semantic segmentation tasks using PyTorch.

📋 Overview

• Exercise 1: Image Classification on CIFAR-10 using CNN architectures (ResNet18, custom CNN, Vision Transformer)
• Exercise 2: Semantic Segmentation on Cityscapes and Oxford-IIIT Pet datasets using SegFormer and FCN

🚀 Exercise 1: Image Classification

Models: ResNet18, Custom CNN, Vision Transformer
Dataset: CIFAR-10 (60k images, 10 classes)
Features: Data augmentation, regularization, advanced optimizers, accuracy metrics

🎯 Exercise 2: Semantic Segmentation

Models: SegFormer, FCN-ResNet50
Datasets: Oxford-IIIT Pet (3 classes), Cityscapes (19 classes)
Features: mIoU metrics, pre-training, fine-tuning

🛠️ Setup and Installation

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)

Installation

1. Clone the repository:

git clone <repository-url>
cd DLVC

2. Install dependencies for Exercise 1:

cd exercise1
pip install -r requirements.txt

3. Download datasets:
   • CIFAR-10: Download from official website
   • Cityscapes: Contact course instructors for preprocessed subset
   • Oxford-IIIT Pet: Automatically downloaded via torchvision

🏃‍♂️ Usage

Exercise 1: Image Classification

Train ResNet18:

cd exercise1
python train_resnet18.py

Train custom CNN:

python train_yourCNN.py

Train Vision Transformer:

python train_yourViT.py

Test models:

python test_resnet18.py
python test_yourCNN.py
python test_yourViT.py

Generate result visualizations:

python generate_graphs.py

Exercise 2: Semantic Segmentation

Train SegFormer:

cd exercise2
python train_segformer.py

Train FCN:

python train.py

Visualize results:

python viz_pets.py

📊 Results

Experimental results are stored in exercise1/tested_configs/ and exercise2/training/ with extensive hyperparameter exploration and performance comparisons.

📈 Features

• Weights & Biases / TensorBoard logging
• Comprehensive metrics and visualization
• Configurable training pipelines
• Pre-training and fine-tuning support

📄 License

Educational project for Deep Learning for Visual Computing course.