Navigation auf uzh.ch
Navigation auf uzh.ch
The prediction of facial attributes from face images has lately gained tremendous attention in the literature, for example: MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes, AFFACT: Alignment-Free Facial Attribute Classification Technique, or Improving Facial Attribute Prediction using Semantic Segmentation. One important driver of the new inventions in that topic is the large-scale CelebA dataset, in which facial images are labeled by hand with 40 different facial attributes such as Big Nose, Bushy Eyebrows, Male, etc. While some of these attributes are independent of the person, e.g., Smiling, Wearing Hat, etc., most of the attributes can be bound to the subject identity. Although the dataset also provides identity information for each image, this information has not yet been utilized in order to improve attribute prediction. For example, if the prediction of the Narrow Eyes attribute is positive in 9 images of one subject, there is a very high probability that the prediction should be positive on a tenth image of the same subject.
The proposed Bachelor's thesis consists of several tasks. The first task should identify which of the 40 facial attributes are stable across subject identity. For this purpose, the ground truth labels of the CelebA training set well be exploited in a statistical manner, maybe treating the absence and the presence of attributes differently. For example, while the presence of a Mustache might vary for some subjects, the absence of a Mustache might be very stable across other identities. Afterward, it is analyzed whether the prediction of the attributes by a state-of-the-art attribute classification technique AFFACT: Alignment-Free Facial Attribute Classification Technique follows similar statistics. Finally, using identity information and the computed statistics, predictions of single attributes should be modified, and it will be tested if such a modification improves the classification accuracy.
If time allows, we will go one step further and do not use ground-truth identity information, but use some face clustering techniques, e.g. ECLIPSE: Ensembles of Centroids Leveraging Iteratively Processed Spatial Eclipse Clustering to cluster samples that belong to the same identity and perform attribute modifications based on the clustered subjects.
