Features and fusion for expression recognition — A comparative analysis

This paper looks at various low-level features, such as Local Binary Pattern (LBP), Local Phase Quantization (LPQ), Scale Invariant Feature Transform (SIFT) and Discrete Cosine Transform (DCT), for performance comparison in subject independent facial expression recognition setting. We use Soft Vector Quantization (SVQ) to compute image-level descriptors. We also do a performance comparison of various pooling methodologies in SVQ. We later do classification using logistic regression followed by fusing likelihoods from the classifiers with various features to come up with joint decisions. Our analysis on the BU-3DFE show that SIFT and mean pooling outperform other features and pooling strategies and that classifier fusion helps in improving the recognition performance.