Reproduce “QCFace: Image Quality Control for boosting Face Representation & Recognition” 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
Duc-Phuong Doan-Ngo, Thanh-Dang Diep, Thanh Nguyen-Duc, Thanh-Sach LE, Nam Thoai
A new loss function, QCFace, enhances face feature representation and recognition by introducing a hard margin strategy that decouples the optimization of feature direction and magnitude. This approach quantifiably encodes image recognizability into feature magnitudes, leading to improved verification accuracy by up to 2.781% on XQLFW and identification accuracy by 6.746% on TinyFace datasets.
Key insights
- Decoupling the optimization of feature direction and magnitude through a hard margin strategy effectively mitigates mutual overlapping gradients, enhancing optimization stability and clarity of representation.
- A detached, probability-based guidance factor enables quantifiable encoding of image recognizability directly into feature magnitudes, creating a broad and distinct magnitude distribution.
- Stable class proxies, such as those provided by ArcFace, are crucial for optimal performance when integrating a recognizability guidance mechanism, as proxy oscillation can degrade results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.