kikogazda/TwinCar-196

🚗 TwinCar: Fine-Grained Car Classification (Stanford Cars 196)

This model predicts car make and model from images using a fine-tuned ResNet50 architecture, trained on the Stanford Cars 196 dataset.

Final project for Brainster Data Science Academy 2025.

---

📝 Project Overview

• Task: Classify car images into 196 make/model categories
• Context: Automated vehicle inspection for TwinCar (drones/robots)
• Input: RGB image (JPG/PNG, 224x224 or similar)
• Output: Predicted car make/model (e.g., "BMW X5")
• Production goal: Identify car brand/model for real-world inspection/automation

---

🏗️ Model Details

• Architecture: ResNet50 (PyTorch, transfer learning, custom head)
• Classes: 196 (make/model, see test_labels.csv)
• Dataset: Stanford Cars 196
• Training: 15 epochs, batch size 32, strong data augmentation
• Authors: Kiril Mickovski & Team 3 (Brainster DSA)
• License: MIT

---

Metrics & Results

Metric	Value	Description
Validation Accuracy	40.6%	Correct predictions on the validation dataset
Training Accuracy	74.4%	Correct predictions on the training dataset
Validation Loss	3.07	Cross-entropy loss on validation set
Training Loss	1.74	Cross-entropy loss on training set
Macro F1-score	0.38	Harmonic mean of precision and recall across all classes
Macro Precision	0.44	Average precision over all classes
Macro Recall	0.41	Average recall over all classes

Full per-class metrics: classification_report.csv

---

📊 Model Evaluation & Explainability

Below are selected screenshots and artifacts showcasing model evaluation, predictions, and explainability:

---

🔵 Confusion Matrix

---

🟢 Training & Validation Curves

---

🟣 F1, Precision, Recall Curves

---

🟠 Top-20 Most Accurate Classes

---

🟡 Top-20 Most Confused Classes

---

🔴 Grad-CAM++ Examples (Explainability)

Model attention visualizations for interpretability (what parts of the image the model “looks at”):

---

📂 Files & Artifacts

• resnet50_finetuned.pth — Model weights
• classification_report.csv — Per-class metrics
• test_predictions_named.csv — Sample predictions
• requirements.txt — Dependencies for inference
• All PNGs — Visualizations, explainability, confusion matrix, curves, etc.

---

🛠️ How to Reproduce

1. Clone this repo or download weights and artifacts from above.
2. Install dependencies:

   pip install -r requirements.txt
   

3. Run the usage code (see Usage section above) or visit the GitHub repo for end-to-end training and evaluation.

---

⚠️ Limitations

• Only 196 classes covered (see Stanford Cars dataset)
• Performance may drop on night images, occlusions, or cars outside the 2010–2012 range
• Trained for car make/model only (not year)

---

👥 Contributors

• Kiril Mickovski
• Team 3, Brainster Data Science Academy 2025

---

📜 Citation

Mickovski, K., Team 3 (2025). TwinCar: Fine-Grained Car Classification (Stanford Cars 196). Brainster Data Science Academy.

---

🔗 Resources

• Stanford Cars 196 Dataset
• GitHub Repo (full code, notebooks)
• Hugging Face Demo Space

---

🧑‍💻 Usage (PyTorch)

import torch
from torchvision import models, transforms
from PIL import Image

model = models.resnet50()
model.fc = torch.nn.Linear(model.fc.in_features, 196)
model.load_state_dict(torch.load("resnet50_finetuned.pth", map_location="cpu"))
model.eval()

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
img = Image.open("your_image.jpg").convert("RGB")
input_tensor = transform(img).unsqueeze(0)

with torch.no_grad():
    logits = model(input_tensor)
    pred = logits.argmax(1).item()