Robust face tracking via collaboration of generic and specific models

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Object tracking, particularly faces, is challenged by significant appearance changes due to pose and expression variations, leading to tracking failures.
• Existing methods often struggle with large pose variations and online adaptation of appearance models.

Purpose of the Study:

• To develop a robust collaborative tracking framework for faces that can handle large pose and expression changes.
• To enable online learning of specific face appearance models for improved tracking accuracy.

Main Methods:

• A dynamic Bayesian network probabilistically combines an offline-trained generic face model with online-learned specific face appearance models.
• A mixture of probabilistic principal component analysis (MPPCA) model represents specific face appearance across multiple views.
• An online Expectation-Maximization (EM) algorithm incrementally updates the MPPCA model using tracking results.

Main Results:

• The collaborative tracking framework demonstrates robustness against large pose and expression variations.
• Online learning of specific appearance models enhances tracking performance for individual faces.
• The method effectively handles distractions from the background during tracking.

Conclusions:

• The proposed collaborative tracking framework offers a robust solution for multiview face tracking under challenging conditions.
• Online learning of appearance models is crucial for adapting to dynamic changes in face appearance.
• This approach significantly improves the reliability of face tracking systems in real-world scenarios.