👦🏻 Auto Face Pose Capture using Logistic Regression

🚀 Problem Statement

In our computer vision application, we wanted to add a person’s face data into a database by capturing it from a camera. Typically, this requires capturing images from multiple facial angles manually — such as:

• Left profile
• Right profile
• Upward (top) view
• Downward (bottom) view
• Center (frontal) view

Manually clicking images while instructing a person to rotate their head is time-consuming and error-prone.

🎯 Objective

Automate the face pose capturing process using machine learning. Once the camera starts, the model should automatically identify the current head position and capture the face when it matches any of the five desired directions.

---

🧠 ML Solution

To solve this, we used InsightFace's face pose estimation features — extracting 3 key pose parameters:

• yaw
• roll
• pitch

Using these, we created a dataset where each face image was labeled based on its pose direction. Then, we trained a Logistic Regression model using a total of 150 features derived from the face landmarks and pose data.

---

📊 Model Training & Evaluation

• ✅ Model: Logistic Regression
• 📈 Accuracy: ~90%
• 🔢 Features Used: 150 pose-related features
• 🧪 Evaluation: Based on test set performance and live camera testing.

---

📦 Project Structure

 ├──app
    ├── captured_face_sides/            # all captured face images
    ├── configs/        # all configuration
    ├── data/           # all data related scripts
    ├── models/      
    ├── src/               
    ├── ..
    ├── .. 
    ├── live.py
    ├── main.py
           
└── README.md            # Project documentation

---

🎥 Demo Video

Watch the project demo here: Demo Video on Google Drive

🎥 Real-time Demo

Run the following to test the model on a live webcam feed:

git clone https://github.com/ak-123459/LookWhere.git

cd LookWhere

pip install -r requirements.txt

python app/live.py

Train and validate the model

python app/main.py

🧠 Model

Algorithm: Logistic Regression

Input: Pose features (yaw, pitch, roll, etc.)

Output: Face direction label

📂 Configuration

Hyperparameters and model configs can be found and edited in config.yaml.

📃 License

This project is licensed under the MIT License.

👤 Developed by: Akash Prasad Mishra

🏢 Parent Organization: Positive Pairs