Navigation auf uzh.ch
Navigation auf uzh.ch
The task of facial attribute classification is to assign several facial attributes (40 in the case of the CelebA Dataset that we will be using) -- such as Oval Face, Narrow Eyes, Bald, Male -- to a given face image. Before the invention of the AFFACT network, facial attribute classification highly relied on face alignment, i.e., facial landmarks had to be detected in order to put the face upright and at a specific location inside of the image. By training our AFFACT network to be more robust to face misalignment, we showed that we can omit the error-prone landmark localization step and rely on the face detector only.
We have also shown in AFFACT that we can make use of the fact that the network is agnostic to misalignment and perform test data augmentation by deliberately changing the scaling, rotation and location of the face in the test image. We extracted attributes for each of these modifications of the image and computed the average as our final prediction. At the end, we found that this process yielded improvements for some of the attributes, but others deteriorated.
The task in this Master's thesis is to find the optimal combination of image transformations for each facial attribute. Since the full span of possible combinations is far too large to be searched, approximative search algorithms such as Evolutionary Algorithms -- which are inspired from how natural evolution works -- should be explored. In an Evolutionary Algorithm, an initial population of random sets of transformations is created, and each of these sets is evaluated. The best-performing sets will assemble the parents, from which pairs are sampled and merged via recombination operations to build new children. Afterward, each child is mutated by randomly changing some of its properties. These mutated children now form the new population, and the process is repeated until the children do not perform better than the parents.
An important part of the Master's thesis is the design of a good representation of the set of transformations and viable recombination and mutation algorithms. Using the final sets of transformations per face, the goal is to improve the state of the art in facial attribute prediction on the CelebA dataset.
