Facial expression recognition in the wild : from individual to group

Abstract

The progress in computing technology has increased the demand for smart systems capable of understanding human affect and emotional manifestations. One of the crucial factors in designing systems equipped with such intelligence is to have accurate automatic Facial Expression Recognition (FER) methods. In computer vision, automatic facial expression analysis is an active field of research for over two decades now. However, there are still a lot of questions unanswered. The research presented in this thesis attempts to address some of the key issues of FER in challenging conditions mentioned as follows: 1) creating a facial expressions database representing real-world conditions; 2) devising Head Pose Normalisation (HPN) methods which are independent of facial parts location; 3) creating automatic methods for the analysis of mood of group of people. The central hypothesis of the thesis is that extracting close to real-world data from movies and performing facial expression analysis on movies is a stepping stone in the direction of moving the analysis of faces towards real-world, unconstrained condition. A temporal facial expressions database, Acted Facial Expressions in the Wild (AFEW) is proposed. The database is constructed and labelled using a semi-automatic process based on closed caption subtitle based keyword search. Currently, AFEW is the largest facial expressions database representing challenging conditions available to the research community. For providing a common platform to researchers in order to evaluate and extend their state-of-the-art FER methods, the first Emotion Recognition in the Wild (EmotiW) challenge based on AFEW is proposed. An image-only based facial expressions database Static Facial Expressions In The Wild (SFEW) extracted from AFEW is proposed. Furthermore, the thesis focuses on HPN for real-world images. Earlier methods were based on fiducial points. However, as fiducial points detection is an open problem for real-world images, HPN can be error-prone. A HPN method based on response maps generated from part-detectors is proposed. The proposed shape-constrained method does not require fiducial points and head pose information, which makes it suitable for real-world images. Data from movies and the internet, representing real-world conditions poses another major challenge of the presence of multiple subjects to the research community. This defines another focus of this thesis where a novel approach for modeling the perception of mood of a group of people in an image is presented. A new database is constructed from Flickr based on keywords related to social events. Three models are proposed: averaging based Group Expression Model (GEM), Weighted Group Expression Model (GEM_w) and Augmented Group Expression Model (GEM_LDA). GEM_w is based on social contextual attributes, which are used as weights on each person's contribution towards the overall group's mood. Further, GEM_LDA is based on topic model and feature augmentation. The proposed framework is applied to applications of group candid shot selection and event summarisation. The application of Structural SIMilarity (SSIM) index metric is explored for finding similar facial expressions. The proposed framework is applied to the problem of creating image albums based on facial expressions, finding corresponding expressions for training facial performance transfer algorithms. Show more