Date: January 2023.
Source: IEEE International Conference on Automatic Face and Gesture Recognition (FG), pp 1–6.
Research Summary: Existing 4D facial expression datasets are limited to a neutral-to-peak-to-neutral transition, but real facial behavior continuously shifts between expressions. Florence 4D addresses this gap with a dataset of dynamic 3D facial expression sequences combining real-world and synthetic identities, all sharing the same mesh topology as the CoMA dataset. The dataset includes 95 identities: 12 real-world subjects from CoMA, itself built from real-world 3dMD scans, 20 subjects captured on a 3dMD scanner, and 63 synthetic identities. These are animated across 70 expressions built from combinations of eight primary emotion categories. Unlike prior datasets, Florence 4D includes expression-to-expression transitions, not just neutral-to-peak sequences, with randomized transition timing for improved realism, producing over 205,000 sequences in total.
3dMD’s Role: Florence 4D draws on real-world 3dMD data from two separate sources. The team’s own real-world 3D scans, 20 subjects, 5 female and 15 male, were captured directly with a 3dMD HR scanner. Separately, Florence 4D incorporates all 12 real-world identities from CoMA, whose own real-world training data was captured on a 3dMD active stereophotogrammetry system at 60 FPS. Together, these two sources anchor the dataset in real-world data rather than relying entirely on synthetic identities.
Article: The Florence 4D Facial Expression Dataset.
Authors: Filippo Principi, Stefano Berretti, Claudio Ferrari, Naima Otberdout, Mohamed Daoudi, Alberto Del Bimbo. Media Integration and Communication Center, University of Florence, Italy and University Lille, France.