A scalable formulation of probabilistic linear discriminant analysis: applied to face recognition

Area of Science:

• Machine Learning
• Computer Vision
• Biometrics

Background:

• Probabilistic Linear Discriminant Analysis (PLDA) is a powerful model for face and speaker recognition.
• Traditional PLDA implementations face scalability issues due to large matrix computations during training.

Purpose of the Study:

• To present a scalable and exact solution for PLDA.
• To overcome the computational limitations of existing PLDA methods.
• To enable improved performance through the use of larger training datasets.

Main Methods:

• Developed a novel, scalable derivation for PLDA that is theoretically equivalent to exact solutions.
• Avoided the need for variational approximations, maintaining accuracy.
• Implemented and tested the scalable PLDA formulation on benchmark datasets.

Main Results:

• Demonstrated the scalability of the proposed PLDA method.
• Validated the equivalence of the scalable implementation with previous non-scalable exact solutions.
• Showcased performance gains on large datasets (Multi-PIE) by utilizing more training samples per identity.

Conclusions:

• The proposed scalable PLDA formulation offers an exact and efficient alternative to existing methods.
• This advancement facilitates the application of PLDA to larger datasets and more complex recognition tasks.
• The method enhances recognition performance by allowing for more comprehensive training data utilization.