GBU-Palm: Multimodal Video Dataset for Palm PAD
A new large-scale multimodal video dataset and benchmark, named GBU-Palm, has been developed for detecting palm presentation attacks (PAD). This dataset consists of 21,326 videos featuring 105 subjects and 210 palms, recorded in six different environments. It encompasses genuine, Print, and Replay presentations, including 6,310 synchronized RGB-NIR samples. The benchmark employs leakage-controlled protocols to distinguish between palm identity and attack lineage, assessing four key video architectures in both environment-matched and held-out scenarios. Findings reveal considerable architecture-dependent performance decline due to environmental changes, and RGB-NIR fusion does not consistently surpass RGB-only input. The research also examines model performance through true accept (TA) and true reject rates, addressing the shortcomings of existing static or limited PAD datasets.
Key facts
- Dataset contains 21,326 videos from 105 subjects and 210 palms.
- Videos captured across six acquisition environments.
- Includes bona fide, Print, and Replay presentation attacks.
- 6,310 synchronized RGB-NIR samples included.
- Leakage-controlled protocols separate palm identity and attack lineage.
- Four video architectures benchmarked under environment-matched and held-out-environment settings.
- RGB-NIR fusion does not consistently outperform RGB-only input.
- Results show architecture-dependent degradation under environmental shift.
Entities
—