A Leopard Cannot Change Its Spots: Improving Face Recognition Using 3D-based Caricatures

• Autor(a) principal: Neves, João
• Data de Publicação: 2018
• Outros Autores: Proença, H.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.6/9171
• topic: Face recognition; 3D caricature

Caricatures refer to a representation of aperson in which the distinctive features are deliberatelyexaggerated, with several studies showing that humansperform better at recognizing people from caricaturesthan using original images. Inspired by this observa-tion, this paper introduces the first fully automatedcaricature-based face recognition approach capable ofworking with data acquired in the wild. Our approachleverages the 3D face structure from a single 2D imageand compares it to a reference model for obtaininga compact representation of face features deviations.This descriptor is subsequently deformed using a ’mea-sure locally, weight globally’ strategy to resemble thecaricature drawing process. The deformed deviationsare incorporated in the 3D model using the Laplacianmesh deformation algorithm, and the 2D face cari-cature image is obtained by projecting the deformedmodel in the original camera-view. To demonstratethe advantages of caricature-based face recognition, wetrain the VGG-Face network from scratch using eitheroriginal face images (baseline) or caricatured images,and use these models for extracting face descriptorsfrom the LFW, IJB-A and MegaFace datasets. The ex-periments show an increase in the recognition accuracywhen using caricatures rather than original images.Moreover, our approach achieves competitive resultswith state-of-the-art face recognition methods, evenwithout explicitly tuning the network for any of theevaluation sets.