Automatic Detection and Intensity Estimation of Spontaneous Smiles
Jeffrey M. Girard · D-Scholarship@Pitt (University of Pittsburgh) · 2014
Both the occurrence and intensity of facial expression are critical to what the face reveals. While much progress has been made towards the automatic detection of expression occurrence, controversy exists about how best to estimate expression intensity. Broadly, one approach is to adapt classifiers trained on binary ground truth to estimate expression intensity. An alternative approach is to explicitly train classifiers for the estimation of expression intensity. We investigated this issue by comparing multiple methods for binary smile detection and smile intensity estimation using two large databases of spontaneous expressions. SIFT and Gabor were used for feature extraction; Laplacian Eigenmap and PCA were used for dimensionality reduction; and binary SVM margins, multiclass SVMs, and ε-SVR models were used for prediction. Both multiclass SVMs and ε-SVR classifiers explicitly trained on intensity ground truth outperformed binary SVM margins for smile intensity estimation. A surprising finding was that multiclass SVMs also outperformed binary SVM margins on binary smile detection. This suggests that training on intensity ground truth is worthwhile even for binary expression detection.