Head-Pose-Estimation

---

This work is done with collaboration of my colleague and my friend: Dina Zakaria

Description

---

This project was developed for angle/pose estimation (pitch, yaw, roll) using 68 landmarks from DLib with models trained on AFLAW2000 Dataset

💡 3D Reconstruction from 2D reference was performed.

Data Preprocessing

---

After reading the data

1. Normalization of all the landmarks to one referenced landmark.

   

   • Using the sparse representation of the normalized coordinates in this paper outline 3.1
   • Given that we have the nose landmark [31] performing as the CG feature of the face
     • We have used it instead of the approximation performed in the paper to estimate its position
     • Then we applied the normalization method using the nose as our CG.

2. Feature Selection

3. Using the Interquartile Rule to Find Outliers in both (labels and landmarks)

   • Calculate the interquartile range for the data. Multiply the interquartile range (IQR) by 1.5  (a constant used to discern outliers). Add 1.5 x (IQR) to the third quartile. Any number greater than this is a suspected outlier.

4. Normalization of all selected features from [0, 1]

Model

---

1. Dividing data using train_test_split from sklearn.model_selection
   • Training: 60%
   • Validation: 20%
   • Testing: 20%
2. Histogram plot of all training, validation and testing data to make sure they all follow a gaussian distribution.
3. Training using six models:
   • Linear Regression
   • Decision tree Regressor
   • Support vector Regression (SVR)
     • Grid Search
   • RandomForestRegressor
   • GradientBoostingRegressor
   • AdaBoostRegressor

Model Performance and Landmarks

---

• Pitch: 12 landmarks were used
  • R2-score for testing data: 92.12% (SVR using Grid Search)
• Yaw: 5 landmarks were used
  • R2-score for testing data: 99.35% (Random Forest)
• Roll: 5 landmarks were used
  • R2-score for testing data: 93.28% (SVR using Grid Search)

Evaluation Metrices

---

• R2 Score
• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• Mean Absolute Percentage Error (MAPE)
• Explained Variance Score
• Max Error
• Median Absolute Error

Video Trial

---

<img src="(" style="width: 650px; max-width: 100%; height: auto" title="Click to enlarge picture" />

References

• Main paper: Sparse Bayesian Regression for Head Pose Estimation