Car Model Classification Using Deep Learning

Project Overview

A comprehensive deep learning project exploring different approaches to classify car models from images. This project implements and compares three distinct methodologies for fine-grained visual classification of 196 different car models. The project demonstrates:

• Advanced transfer learning techniques with EfficientNetV2L
• Novel application of KNN with deep embeddings
• Custom implementation of ResNet architecture
• Systematic experimentation with model architectures and hyperparameters
• Comprehensive evaluation using multiple performance metrics

Table of Contents

• Environment Setup
• Dataset
• Project Structure
• Methodologies
• Results
• Technical Details
• Future Work
• Contributing
• License

Environment Setup

Google Colab Environment

This project is designed to run in Google Colab, leveraging its GPU capabilities for efficient model training. The notebook includes all necessary setup steps:

1. Google Drive mounting for data persistence
2. Directory structure creation
3. Required package installation
4. GPU runtime configuration

Dataset

The Stanford Cars Dataset provides a challenging benchmark for fine-grained visual classification:

• 196 unique car classes (make, model, year)
• 8,152 training images (split into training and validation)
• 8,048 test images
• Diverse image conditions:
  • Various lighting conditions
  • Different angles and perspectives
  • Complex backgrounds
  • Multiple scales and resolutions

Project Structure

The project is organized to maintain clear separation between data, models, and documentation:

Google Drive/
└── [Project Directory]/
    ├── dataset/                  # Image data
    ├── models/                   # Saved models
    ├── final_project.ipynb      # Main notebook
    └── annotations.xlsx         # Class metadata

dataset/ and models/ directories can be created using designated notebook cells

Methodologies

1. Transfer Learning with EfficientNetV2L

Exploration of transfer learning approaches using EfficientNetV2L as the backbone:

Three Strategic Configurations:

1. Base Architecture

   • Minimalist approach with single dense layer
   • Moderate regularization
   • Baseline for performance comparison

2. Enhanced Capacity

   • Deeper architecture with multiple dense layers
   • Hierarchical feature learning
   • Increased model expressiveness

3. High Regularization

   • Aggressive dropout strategy
   • Focus on generalization
   • Overfitting prevention

2. KNN with Deep Embeddings

An innovative hybrid approach combining deep learning feature extraction with traditional machine learning:

• Feature extraction using EfficientNetV2L
• Exploration of neighborhood dynamics:
  • Tight neighborhoods (k=3) for precision
  • Balanced approach (k=5) for general cases
  • Broad context (k=10) for robustness

3. Custom ResNet Architecture

Implementation of residual networks with systematic depth variation:

• Architecture Philosophy:

  • Skip connections for gradient flow
  • Batch normalization for training stability
  • Strategic depth increases

• Depth Configurations:

  1. Shallow (3 blocks): Fast and efficient
  2. Medium (5 blocks): Balanced complexity
  3. Deep (7 blocks): Maximum feature hierarchy

Technical Details

Data Preprocessing Strategy

Comprehensive image preprocessing pipeline:

• Standardized sizing (224x224)
• Advanced augmentation techniques
• Normalization and standardization
• Efficient data loading and caching

Training Approach

Carefully crafted training configuration:

• Optimized batch sizes for GPU memory
• Strategic learning rate management
• Multi-metric evaluation
• Performance monitoring and validation

Performance Optimization

Focus on efficient resource utilization:

• GPU memory management
• Data pipeline optimization
• Strategic model checkpointing
• Training time optimization

Results

[Coming soon: Comprehensive comparison of model performances, including:

• Accuracy metrics across architectures
• Training dynamics analysis
• Resource utilization comparison
• Error analysis and insights]

Future Work

Potential areas for exploration and improvement:

• Advanced data augmentation strategies
• Model ensemble techniques
• Architecture optimization
• Performance enhancement methods

Author

Idan David Or Lavi Senior Data Scientist

Acknowledgments

• Stanford Cars Dataset creators
• TensorFlow and Google Colab teams
• Kaggle community

Hardware Requirements

• GPU: Google Colab (Tesla T4/P100)
• RAM: 12GB+ (Colab provided)
• Storage: Google Drive space for dataset

Runtime Notes

• The notebook is optimized for Google Colab's GPU runtime
• Training times may vary based on Colab's GPU availability
• Consider using Colab Pro for more consistent GPU access