Reproduce “Multi-View Slot Attention Using Paraphrased Texts for Face Anti-Spoofing” 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
Jeongmin Yu, Susang Kim, Kisu Lee, Taekyoung Kwon, Won-Yong Shin, Ha Young Kim
Researchers at Yonsei University developed MVP-FAS, a framework that enhances Face Anti-Spoofing generalization by leveraging CLIP's patch embeddings with multi-view slot attention and aligning them with paraphrased texts. It consistently achieved state-of-the-art cross-domain performance, reducing the average HTER by 0.8 percentage points on Protocol 1 and 2.74 percentage points on Protocol 2.
Key insights
- Directly leveraging CLIP's fine-grained patch embeddings, enriched with global context, is crucial for detecting subtle local spoofing artifacts and outperforms global-feature-only approaches.
- Employing multiple, paraphrased textual descriptions for 'live' and 'spoof' classes significantly enhances semantic robustness and model generalization compared to single-text prompts.
- The synergistic combination of multi-view slot attention (MVS) and explicit patch-level semantic alignment (MTPA) drastically improves cross-domain FAS performance and model interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.