Towards Facial Emotion Recognition using a Transformer-based Model with 3D Facial Pointclouds

Facial Emotion Recognition (FER) stands as a pivotal area of research within Affective Computing (AAC) due to its broad applicability in domains such as human-robot interaction (HRI), mental health assessment, and fatigue monitoring. Current state-of-the-art FER approaches are based on deep learning...

Descripción completa

Detalles Bibliográficos
Autor: Rayón Ropero, Laura
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/422399
Acceso en línea:https://hdl.handle.net/2117/422399
Access Level:acceso abierto
Palabra clave:Human face recognition (Computer science)
Transfer learning (Machine learning)
Deep learning (Machine learning)
3D Facial Emotion Recognition
Joint Communication and Sensing (JC&S)
Transformers
Deep Learning
Transfer Learning
Reconeixement facial (Informàtica)
Aprenentatge automàtic
Aprenentatge profund
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Reconeixement de formes
Descripción
Sumario:Facial Emotion Recognition (FER) stands as a pivotal area of research within Affective Computing (AAC) due to its broad applicability in domains such as human-robot interaction (HRI), mental health assessment, and fatigue monitoring. Current state-of-the-art FER approaches are based on deep learning methods and primarily leverage 2D image data. However, the acquisition of such data through off-person sensors poses significant privacy challenges under EU regulations. In response, we envision a trend shift towards on-person sensors integrated into wearables, facilitated by the Joint Communication and Sensing (JC&S) paradigm. This paradigm proposes the use of high-frequency electromagnetic waves for fine-grained structural 3D imaging of the human face, enabling continuous and privacy-preserving emotion tracking. To support this emerging paradigm, this work addresses the two open challenges facing 3D-based FER: the limited size of 3D databases labelled for FER, and the need for understanding the use cases where Transfer Learning for 3D-based DL models is applicable. To overcome the first challenge, we showcase the potential of using a previously developed in-house method for generating 3D FER databases from existing 2D image FER databases. Obtained through this method, we use a 3D version of the popular 2D FER AffectNet database for 3D FER DL model training, achieving a 31.62% accuracy. Furthermore we introduce a data refining pipeline to facilitate the model's learning process. By isolating the facial region of the 3D point clouds, this refining process allows further boosting classification accuracy by 12%, which confirms the potential of our approach. To overcome the second challenge, we pioneer the use of transfer learning from non-FER 3D domains to FER-3D domains. We fine-tune the transformer-based DL model I2P-MAE, pre-trained on the ShapeNet database of CAD objects, and achieve the promising result of 90.12% accuracy on the well-established 3D FER BU-3DFE dataset.