Identity Level Attributes for Explainable Facial Verification
Loading...
Authors
Flourens, Cooper
Issue Date
2024
Type
Thesis
Language
Keywords
Attributes , Explainability , Face Verification , Interpretability
Alternative Title
Abstract
Attributes are describable features of faces and are the foundational visual component that humans use to recognize faces. Current state-of-the-art facial recognition methods are end-to-end deep learning systems that are not interpretable by humans. In this thesis, we introduce a new type of facial attribute label representing identity-level prominent features on an existing face image dataset, CarVer, evaluating its performance against another facial attribute dataset, CelebA. Additionally, we expand CarVer with a set of images from internet sources, creating a new dataset we call CarVerX. This thesis analyzes the utility of identity-level prominent features for explainable face verification. We find that our prominent features outperform traditional facial attributes for the task of face verification opening the door for future research along this avenue for explainable face verification.