Date of Award

December 2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Zeyun Yu

Abstract

Facial landmark detection on 3D human faces has had numerous applications in the literature

such as establishing point-to-point correspondence between 3D face models which is itself a

key step for a wide range of applications like 3D face detection and authentication, matching,

reconstruction, and retrieval, to name a few.

Two groups of approaches, namely knowledge-driven and data-driven approaches, have been

employed for facial landmarking in the literature. Knowledge-driven techniques are the

traditional approaches that have been widely used to locate landmarks on human faces. In

these approaches, a user with sucient knowledge and experience usually denes features to

be extracted as the landmarks. Data-driven techniques, on the other hand, take advantage

of machine learning algorithms to detect prominent features on 3D face models. Besides

the key advantages, each category of these techniques has limitations that prevent it from

generating the most reliable results.

In this work we propose to combine the strengths of the two approaches to detect facial

landmarks in a more ecient and precise way. The suggested approach consists of two phases.

First, some salient features of the faces are extracted using expert systems. Afterwards,

these points are used as the initial control points in the well-known Thin Plate Spline (TPS)

technique to deform the input face towards a reference face model. Second, by exploring and

utilizing multiple machine learning algorithms another group of landmarks are extracted.

The data-driven landmark detection step is performed in a supervised manner providing an

information-rich set of training data in which a set of local descriptors are computed and used

to train the algorithm. We then, use the detected landmarks for establishing point-to-point

correspondence between the 3D human faces mainly using an improved version of Iterative

Closest Point (ICP) algorithms. Furthermore, we propose to use the detected landmarks for

3D face matching applications.

Share

COinS