Harnessing Generated Instance-Specific Neutral Expressions in Facial Expression Classification

摘要

How can we harness neutral expressions in facial expression classification, in the absence of explicit provided neutral state of each face? To address this question, we chiefly introduce a novel conditional generative model, e.g., the conditional encapsulated variational auto-encoders (EVAE), to generate our targeted instance-specific neutral expressions. Running our introduced conditional EVAE on the FER2013 dataset, it exhibits superior generative performance on generating instance-specific neutral expressions, in the face of other generative models. Armed with generated instance-specific neutral expressions, we put forth two applications of these neutral expressions, ranging from the proposal of the proactive label relaxation technique to the introduction of a novel multi-task learning model. Experimental results reflect the effectiveness of these two applications in tackling facial expression classification.