Reproduce “IFAST: Weakly Supervised Interpretable Face Anti-spoofing from Single-shot Binocular NIR Images” on OpenResearch

About this paper

Jiancheng Huang, Donghao Zhou, Shifeng Chen

Researchers at the Shenzhen Institute of Advanced Technology developed IFAST, a weakly supervised interpretable face anti-spoofing method utilizing single-shot binocular Near-Infrared (NIR) images. The work also introduces BNI-FAS, the first large-scale binocular NIR dataset, enabling IFAST to achieve 99.41% accuracy and 0.3657% Equal Error Rate while generating pixel-wise disparity maps for decision interpretability without extensive manual annotations.

Key insights

• Binocular Near-Infrared (NIR) imaging offers inherent advantages for Face Anti-Spoofing, such as immunity to screen-based attacks and robustness in low-light environments.
• Effective disparity estimation for FAS can be achieved with weak supervision by leveraging a dual-teacher distillation framework to generate pseudo-depth maps, significantly reducing annotation costs.
• A specially adapted Transformer architecture with Dynamic Matching Attention efficiently processes binocular NIR images to produce accurate and interpretable disparity maps, providing explicit evidence for anti-spoofing decisions.

Summary by alphaXiv.