Human identification is getting ubiquitous in our everyday lives, where algorithms verify identities based on various physical and behavioural attributes. This thesis is focusing on one step in this process, that is detecting facial landmarks, which is usually the first step in the identification process, on several image databases. Specifically, it compares three detection models, which are based on computer vision. Those are Supervised Descent Method or SDM, Tasks-constrained deep convolutional network or TCDCN for short and Multi-Center Network, also called MCNet. The goal of the thesis is to understand how these methods behave under varying circumstances, how image and facial characteristics influence their success and to determine their respective advantages and disadvantages.
Firstly, we lay the foundation with an overview of existing research, that provides a broader view of the state of the profession. In the theoretical part is then presented the basic theoretical background of detecting facial landmarks and the chosen methods for their detection that will be evaluated. The section on the methodology of the work describes the image databases used in the experiments and the methodology for evaluating the results. Installation and implementation of the methods and the tools used are also described in depth. As part of the thesis, 5 experiments were carried out, which focused on the verification of methods for certain categories of images according to gender, race and proximity of persons, as well as according to the importance of the colour information of the images and the conditions they were captured under. Lastly, the results and the problems encountered by the methods in detecting facial landmarks were presented, as well as general observations and findings of the thesis. The achieved results are briefly summarised and suggestions for further research are given.
|
