Face Verification with Veridical and Caricatured Images using Prominent Attributes
Loading...
Authors
Sutariya, Jayam V.
Issue Date
2024
Type
Thesis
Language
Keywords
Caricatures , Explainability , Face Verification , Facial Attributes , Feature Classification
Alternative Title
Abstract
Caricatures, with their exaggerated features, offer a surprisingly efficient means for individuals to recognize each other compared to veridical (real) images. However, it is still a difficult task in machine learning to match veridical images to caricatures. This is due to the poor quality of caricature datasets, which often lack clear labels and contain low-quality images. Widely utilized veridical image datasets like CelebA also suffer from inadequate labeling. These label inconsistencies pose significant issues in accurate face verification tasks. Moreover, the effectiveness of neural networks has led to a shift in focus away from attribute-based representations, further complicating the matching process. In this thesis, we introduce a classification protocol for prominent facial feature recognition along with a verification protocol for matching celebrity veridical images to their caricatures. We utilize CarVer, a recently curated dataset comprising both veridical and caricature images accompanied by detailed prominent feature labels.
Description
Citation
Publisher
License
Creative Commons Attribution-NonCommercial 4.0 International