Doubly Sparse Relevance Vector Machine for Continuous Facial Behavior Estimation.

Area of Science:

• Affective computing
• Machine learning
• Biomedical signal processing

Background:

• Subjective states like pain are difficult to measure directly.
• Facial expressions offer a window into these internal states.
• Analyzing subtle facial movements is crucial for accurate interpretation.

Purpose of the Study:

• To develop a novel regression method for continuous estimation of facial behavior intensity.
• To introduce the Doubly Sparse Relevance Vector Machine (DSRVM) for enhanced facial expression analysis.
• To improve upon existing multi-kernel learning techniques by incorporating kernel sparsity.

Main Methods:

• Formulated a new regression technique: Doubly Sparse Relevance Vector Machine (DSRVM).
• Enforced double sparsity by selecting relevant training examples (relevance vectors) and important kernels.
• Applied multi-kernel learning with a focus on sparsity of relevant kernels.

Main Results:

• DSRVM demonstrated superior performance compared to existing methods.
• Achieved a multi-fold reduction in training and testing running times.
• Empirically validated on Shoulder Pain videos, DISFA, and SEMAINE datasets.

Conclusions:

• DSRVM offers an efficient and accurate approach for estimating facial behavior intensity.
• The method effectively integrates feature and kernel selection for improved interpretation.
• This work advances the field of facial expression analysis, particularly for subjective states like pain.