Use this URL to cite or link to this record in EThOS:
Title: Photorealistic retrieval of occluded facial information using a performance-driven face model
Author: Berisha, F.
ISNI:       0000 0004 2731 5432
Awarding Body: University College London (University of London)
Current Institution: University College London (University of London)
Date of Award: 2009
Availability of Full Text:
Access from EThOS:
Access from Institution:
Facial occlusions can cause both human observers and computer algorithms to fail in a variety of important tasks such as facial action analysis and expression classification. This is because the missing information is not reconstructed accurately enough for the purpose of the task in hand. Most current computer methods that are used to tackle this problem implement complex three-dimensional polygonal face models that are generally timeconsuming to produce and unsuitable for photorealistic reconstruction of missing facial features and behaviour. In this thesis, an image-based approach is adopted to solve the occlusion problem. A dynamic computer model of the face is used to retrieve the occluded facial information from the driver faces. The model consists of a set of orthogonal basis actions obtained by application of principal component analysis (PCA) on image changes and motion fields extracted from a sequence of natural facial motion (Cowe 2003). Examples of occlusion affected facial behaviour can then be projected onto the model to compute coefficients of the basis actions and thus produce photorealistic performance-driven animations. Visual inspection shows that the PCA face model recovers aspects of expressions in those areas occluded in the driver sequence, but the expression is generally muted. To further investigate this finding, a database of test sequences affected by a considerable set of artificial and natural occlusions is created. A number of suitable metrics is developed to measure the accuracy of the reconstructions. Regions of the face that are most important for performance-driven mimicry and that seem to carry the best information about global facial configurations are revealed using Bubbles, thus in effect identifying facial areas that are most sensitive to occlusions. Recovery of occluded facial information is enhanced by applying an appropriate scaling factor to the respective coefficients of the basis actions obtained by PCA. This method improves the reconstruction of the facial actions emanating from the occluded areas of the face. However, due to the fact that PCA produces bases that encode composite, correlated actions, such an enhancement also tends to affect actions in non-occluded areas of the face. To avoid this, more localised controls for facial actions are produced using independent component analysis (ICA). Simple projection of the data onto an ICA model is not viable due to the non-orthogonality of the extracted bases. Thus occlusion-affected mimicry is first generated using the PCA model and then enhanced by accordingly manipulating the independent components that are subsequently extracted from the mimicry. This combination of methods yields significant improvements and results in photorealistic reconstructions of occluded facial actions.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available