Reproduce “Exploring Disentangled Feature Representation Beyond Face Identification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yu Liu, Fangyin Wei, Jing Shao, Lu Sheng, Junjie Yan, Xiaogang Wang
A D²AE framework learns disentangled and complementary face features from identity supervision alone, effectively separating identity-specific traits from identity-agnostic information. This framework achieves 99.80% accuracy on LFW for verification and 87.82% accuracy on CelebA for attribute recognition, demonstrating capabilities in controllable face generation and editing.
Key insights
- Adversarial supervision, specifically an identity confusion loss and an identity classification loss, effectively forces identity-dispelled features to suppress identity information while capturing other facial characteristics.
- Identity-distilled features accurately capture identity-specific traits, forming dense, distinct clusters, while identity-dispelled features excel at recognizing identity-irrelevant attributes.
- Statistical augmentation with Gaussian noise promotes distinctive, densely Gaussian clustered, and nearly independent channels in the latent space, resulting in a compact and convex representation suitable for smooth face editing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.