Identity-invariant Facial Landmark Frontalization for Facial Expression Analysis

We propose a new data-driven frontalization technique for 2D facial landmarks, designed to aid in the analysis of facial expressions. It employs a new normalization strategy aiming to minimize identity variations, by standardizing the relative position of facial parts. It operates directly on 2D landmark coordinates, does not require additional feature extraction and is very `light', adding minimal computational overhead. The proposed approach exhibits considerable improvement over a reference method, justifying its use as preprocessing for expression analysis systems relying on geometric features.