Reproduce “Primitive-Driven Compositional Forensic Visual Prompting for Open-World 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
Fangling Jiang, Qi Li, Bing Liu, Weining Wang, Quilin Huang, Zhenan Sun, Ming-Hsuan Yang
Researchers from CASIA, University of South China, and UC Merced developed Primitive-Driven Compositional Forensic Visual Prompting, a purely visual framework for open-world face anti-spoofing. This approach learns and dynamically composes micro-forensic visual primitives from a frozen Vision Transformer to capture fine-grained spoofing evidence, achieving state-of-the-art performance with a 29.51% relative HTER reduction on the SiW-Mv2 (W) target domain.
Key insights
- Many unseen attacks arise from new combinations of recurring elementary visual forensic cues, enabling a compositional approach to generalization.
- Purely visual prompts, operating in the feature space of a frozen Vision Transformer, can capture superior fine-grained and high-frequency forensic evidence compared to language-mediated prompts.
- Dynamic, input-adaptive composition of micro-forensic primitives, guided by global context, is effective for detecting heterogeneous unseen attacks without relying on pre-defined categories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.