Rehearsal-Free Domain Continual Face Anti-Spoofing: Generalize More and Forget Less, accepted in ICCV 2023 (Oral)
Previous work of domain continual learning for FAS:
- Use rehearsal data to reduce forgetting
- Only evaluate seen data performance (intra-domain)
Our work:
- Use no rehearsal-data based on privacy concerns
- Evaluate models on unseen domain performance
- Cross/Unseen-domain performance is important but ignored in domain continual learning for Face Anti-Spoofing (DCL-FAS)
- The more generalized the model, the less forgetting the model
- We handle cross-domain and catastrophic forgetting performance at the same time
-
We propose Dynamic Central Convolutional Adapter (DCDCA) for strong domain generalized capability
-
We propose Proxy Prototype Contrastive Regularization (PPCR) to reduce forgetting, without rehearsal data
-
We propose the first rehearsal-free domain continual face anti-spoofing protocols based on 15 datasets, which is practically based on the released years
- Base: Large-scale pretraining (all train data used)
- Continual: Low-shot (100 frames used for fine-tuning)
- Evaluation: both seen& unseen
- Dataset ordered according to released years (2012-2022)
- Protocol 1: old to new
- Protocol 2: new to old
