| Summary: | Face alignment is one of the fundamental steps in a vast number of tasks of high economical and social value, ranging from security to health and entertainment. Despite the attention received from the community for more than 2 decades and the success of cascaded regression based approaches, many challenges were yet to be solved, such as the case of near-profile poses and low resolution faces.
In this thesis, we successfully address a series of such challenges in the area of face alignment and super-resolution, significantly pushing the state-of-the-art by proposing novel deep learning-based architectures specially tailored for fine grained recognition tasks. In summary, we address the following problems: (I) fitting faces found in large poses (Chapter 3), (II) in both 2D and 3D space (Chapter 4), creating in the process (III) the largest in-the-wild large pose 3D face alignment dataset (Chapter 4). While the case of high resolution faces was actively explored in the past, in this thesis we systematically study and address a new challenge: that of (IV) fitting landmarks in very low resolution faces (Chapter 6). While deep learning based approaches achieved remarkable results on a wide variety of tasks, they are usually slow having high computational requirements. As such, in Chapter 5, we propose (V) a novel residual block carefully crafted for binarized neural networks that significantly improves the speed, due to the use of binary operations for both the weights and the activations, while maintaining a similar or competitive accuracy.
The results presented through out this thesis set the new state-of-the-art on both 2D & 3D face alignment and face super-resolution.
|
|---|