Nonparametric Facial Feature Localization Using Segment-Based Eigenfeatures.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Facial feature localization is crucial for many computer vision applications.
• Existing methods often rely on iterative optimization or search algorithms, which can be computationally expensive.
• There is a need for efficient and accurate facial feature localization techniques.

Purpose of the Study:

• To present a novel nonparametric method for facial feature localization.
• To achieve fast and accurate localization without iterative parameter optimization or search algorithms.
• To explore model size reduction techniques for trained models.

Main Methods:

• Utilizes relative directional information between image segments and facial feature points.
• Employs a weighted concentration of directional vectors pointing to feature positions.
• Calculates directional vectors using eigendirectional vectors from Principal Component Analysis (PCA) of Histogram of Oriented Gradient (HOG) features.
• Leverages statistical reasoning from training data.

Main Results:

• Achieves very fast and accurate facial feature localization.
• Avoids the need for local pattern extraction at estimated feature positions.
• Eliminates the requirement for iterative parameter optimization or search algorithms.
• Enables reduction of trained model storage size by controlling HOG pattern space energy preservation.

Conclusions:

• The proposed nonparametric method offers a computationally efficient and accurate solution for facial feature localization.
• The technique's reliance on statistical reasoning and PCA-derived features contributes to its speed and accuracy.
• Model compression is feasible by adjusting the energy preservation level in the HOG feature space.