Sequential emotion recognition using Latent-Dynamic Conditional Neural Fields

A wide number of problems in face and gesture analysis involve the labeling of temporal sequences. In this paper, we introduce a discriminative model for such sequence labeling tasks. This model involves two layers of latent dynamics, each with their separate roles. The first layer, the neural network or gating layer, aims to extract non-linear relationships between input data and output labels. The second layer, the hidden-states layer, aims to model temporal substructure in the sequence by learning hidden-states and their transition dynamics. A new regularization term is proposed for the training of this model, encouraging diversity between hidden-states. We evaluate the performance of this model on an audiovisual dataset of emotion recognition and compare it against other popular methods for sequence labeling.