Abstract

This thesis describes the design and implementation of a unified framework for face
detection and landmark alignment in arbitrary in the wild images. Traditionally, both of
these problems have been addressed separately in literature with impressive results
being recently reported in both of these fields. But, if one was to construct a pipeline
consisting of a state-of-the-art face detection method followed by a state-of-the-art
facial landmark localisation algorithm, the overall performance outcome would not be
proficient enough to be used in high level algorithms such as face recognition and
facial expression. This is because the accuracy produced by the face detector is not
sufficiently high enough to initialise the landmark localisation algorithm.
To address this aforementioned limitation, this thesis aims to propose an approach
that combines both of these tasks into a single unified algorithm that can be run in real
time, by utilising the parallel computing architecture of the graphics processing unit
(GPU). This will be done by using a Cascaded-Regression (CR) algorithm in a sliding
window fashion. The proposed system will exploit the CR algorithms ability to compute
the 2D pose of a face from rough initial estimates, in order to generate a Hough-
Transform voting scheme for detecting candidate faces and filtering out irrelevant
background. The obtained detection surface will then be further refined using SVM to
yield both face detections and the location of their parts.
The proposed system for this thesis will be built within the MATLAB environment, using
a MEX-file which will provide an interface to the proposed CUDA algorithm. The results
of which, will be tested against current state-of-the-art methods for both face detection
and landmark localisation.
We evaluate performance on the most widely used data sets in face detection, namely
annotated faces in-the-wild (AFW) (Zhu and Ramanan, 2012), Face Detection Dataset
and Benchmark (FDDB) (Jain and Learned-Miller, 2010) and Caltech Occluded Faces
in the Wild (COFW) (Burgos-Artizzu, Perona and Dollár, 2013). The empirical results
demonstrate that the proposed unified framework achieves state-of-the-art
performance in both face detection and facial alignment, and that our detector clearly
outperforms all commercial and published methods by a margin of over 10% in
detection accuracy on the AFW dataset.