Car Viewpoint Classifier 🚗

High-accuracy car viewpoint classification for mobile deployment

Live Demos:

• 🚀 Classification Demo → (Trained Model)
• 🔍 FAISS Similarity Demo → (MobileNet Features)

---

📥 Model Download

Pre-trained TFLite Model (Ready for Mobile Deployment)

• Download: viewpoint_classifier_float32.tflite 📦
• Size: 2.40 MB
• Accuracy: 87.22%
• Format: TensorFlow Lite (Float32)
• Input: 224×224×3 RGB images
• Output: 7 classes (6 viewpoints + unknown)

📦 Available Models

Model	Size	Accuracy	Download Link
Float32	2.40 MB	87.22%	Download
INT8	2.59 MB	~85-87%	Download

Class Labels

classes = ["front", "frontleft", "frontright", "rear", "rearleft", "rearright", "unknown"]

---

🚀 Quick Start

import tensorflow as tf
import numpy as np
from PIL import Image

# Load model
interpreter = tf.lite.Interpreter(model_path="viewpoint_classifier_float32.tflite")
interpreter.allocate_tensors()

# Prepare image
image = Image.open("car.jpg").resize((224, 224))
image_array = np.array(image).astype(np.float32) / 255.0
image_array = np.expand_dims(image_array, axis=0)

# Run inference
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.set_tensor(input_details[0]['index'], image_array)
interpreter.invoke()
predictions = interpreter.get_tensor(output_details[0]['index'])[0]

# Get result
classes = ["front", "frontleft", "frontright", "rear", "rearleft", "rearright", "unknown"]
predicted_class = classes[np.argmax(predictions)]
confidence = np.max(predictions)
print(f"Predicted: {predicted_class} ({confidence*100:.2f}%)")

---

📊 Performance

Metric	Value
Overall Accuracy	87.22%
Model Size	2.40 MB
Parameters	2.27M
Architecture	MobileNetV2

Per-Class Accuracy

Viewpoint	Accuracy
Front	96.26% ⭐
Front Right	92.74%
Rear Left	90.45%
Rear Right	89.50%
Rear	83.04%
Unknown	82.57%
Front Left	81.67%

---

🎯 Project Overview

This project implements a 7-class car viewpoint classifier optimized for mobile deployment using TensorFlow Lite.

Viewpoint Classes

1. Front - Front view of the car
2. Front Right - Front-right diagonal view
3. Front Left - Front-left diagonal view
4. Rear - Rear view of the car
5. Rear Right - Rear-right diagonal view
6. Rear Left - Rear-left diagonal view
7. Unknown - Invalid/unrecognized viewpoint

---

📁 Repository Structure

car_side_recog/
├── README.md                          # This file
├── docs/                              # Additional documentation
│   ├── TRAINING_README.md            # Training instructions
│   ├── CLASSIFICATION_RESULTS.md     # Detailed results
│   └── VISUALIZATION_GUIDE.md        # Visualization tools
├── train_viewpoint_classifier_fast.py # Main training script
├── resume_training.py                 # Resume from checkpoint
├── convert_tflite_final.py           # TFLite conversion
├── test_predict.py                    # Inference script
├── requirements.txt                   # Python dependencies
├── class_map.json                     # Class label mapping
└── .gitignore                         # Git ignore rules

---

📚 Documentation

• Training Guide - How to train the model from scratch
• Classification Results - Detailed performance analysis
• Visualization Guide - Dataset visualization tools

---

🛠️ Installation

# Clone the repository
git clone https://github.com/avinash064/Car_side_recog.git
cd Car_side_recog

# Install dependencies
pip install -r requirements.txt

Requirements

• Python 3.9+
• TensorFlow 2.10+
• NumPy
• Pillow
• scikit-learn

---

💻 Training

To train the model from scratch:

# Activate virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Run training
python train_viewpoint_classifier_fast.py

For detailed training instructions, see Training Guide.

---

🔄 TFLite Conversion

Convert your trained model to TFLite:

python convert_tflite_final.py

The converted model will be saved to tflite_models/viewpoint_classifier_float32.tflite.

---

🧪 Testing

Run inference on test images:

python test_predict.py --image path/to/car_image.jpg

---

📝 Model Specifications

• Input Shape: [1, 224, 224, 3]
• Input Type: float32 (normalized 0-1)
• Output Shape: [1, 7]
• Output Type: float32 (softmax probabilities)
• Preprocessing: Resize to 224×224, normalize to [0, 1]

---

🎓 Citation

If you use this model or code in your research, please cite:

@misc{car_viewpoint_classifier_2026,
  author = {Avinash Kashyap},
  title = {Car Viewpoint Classifier},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/avinash064/Car_side_recog}
}

---

📧 Contact

• Author: Avinash Kashyap
• GitHub: @avinash064
• Repository: Car_side_recog

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🙏 Acknowledgments

• MobileNetV2 architecture from TensorFlow
• Training dataset from CFV (Canonical Frontal View) dataset
• TensorFlow Lite for mobile optimization