A Multi-Task Computer Vision Study using EfficientNetB0-based car brand and type classification with bounding boxes
This repository contains a research-oriented computer vision project focused on the simultaneous classification of car brand and vehicle type from images.
The work is implemented entirely as a Jupyter/Colab notebook and emphasizes:
- Model architecture design
- Training and fine-tuning strategy
- Rigorous evaluation and error analysis
The project is suitable for research review, technical interviews, and applied machine learning discussions.
Identifying:
- Car Brand (e.g., Hyundai, Mercedes, BMW)
- Vehicle Type (Sedan, Hatchback, SUV, etc.)
from images is challenging due to:
- High visual similarity across brands
- Background noise and occlusions
- Variations in lighting and viewpoint
This project explores a multi-task learning approach to address these challenges efficiently.
- Backbone: EfficientNetB0 (ImageNet pretrained)
- Design: Multi-output Convolutional Neural Network
- Shared feature extractor
- Separate classification heads for:
- Car Brand
- Vehicle Type
- Transfer learning with frozen backbone
- Progressive fine-tuning of upper layers
- Categorical cross-entropy loss per output
The notebook includes in-depth evaluation, going beyond aggregate accuracy:
- Confusion matrices (raw and normalized)
These analyses provide interpretability and diagnostic insights into model behavior.
- Image metadata handled via CSV files
- Bounding-box guided cropping for vehicle localization
- There are two types of datasets used: (1) train_data and (2) test_data. The folder train_data contains 478 car images with subfolders (a) Convertible, (b) Hatchback and (c) Sedan. The test_data folder contain only 54 car images.
- There is one CSV file which gives a clear information about brands and types of the cars.
- Python
- TensorFlow / Keras
- OpenCV
- NumPy and Pandas
- Scikit-learn
- Matplotlib and Seaborn
- Automotive Analytics
- Intelligent transport systems
- Vehicle inspection platforms
- AI-driven automotive startups
- Research and benchmarking
👉 Check out the link here for more details:
🔗 https://drive.google.com/drive/folders/1qgUuWCbQyLyzA7KEEqHnmphKa1ypvF7F?usp=sharing
MBA (Data Science) -IIM Visakhapatnam
MSc (Physics) - IIT Kharagpur
This project is created and maintained by Parmesh Kumar.
📄 Read more about the author here:
👉 Author.md