Research map / Label-efficient and robust learning / Test-time adaptation
TRAM: Finetuning-Free Test-Time Adaptation for Generalized Face Anti-Spoofing with Only a Few Bonafide Samples
Wang QIRUI, SI-QI LIU
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add I2CL relation modeling to the FLIP-IT baseline (ablation) | TRAM · Protocol 2 (avg) | HTER (%): 3.05 -> 1.70 | Improved |
| add Consensus Attention Regularization (CAR) (ablation) | TRAM (I2CL) · Protocol 2 (avg) | HTER (%): 1.70 -> 1.32 | Improved |
| add Fine-grained Relation Prompt Learning (FRPL) (ablation) | TRAM (I2CL+CAR) · Protocol 2 (avg) | HTER (%): 1.32 -> 0.81 | Improved |
| replace Siamese fusion with standard cross-attention (ablation) | TRAM relation encoder · Protocol 2 (avg) | HTER (%): 3.66 -> 2.80 | Improved |
| add anchor-input difference in query and key (RCA Q/K) (ablation) | cross-attention encoder · Protocol 2 (avg) | HTER (%): 2.80 -> 2.04 | Improved |
| add mask-aware temperature modulation to RCA (ablation) | RCA · Protocol 2 (avg) | HTER (%): 2.04 -> 1.70 | Improved |
| use Q-difference only versus Q/K difference (ablation) | RCA · Protocol 2 (avg) | HTER (%): 2.40 -> 2.04 | Improved |
| use intensity-based instead of semantic anchor selection (ablation) | TRAM · Protocol 2 (avg) | HTER (%): 0.97 -> 0.81 | No clear effect |
| increase number of bonafide anchor shots beyond 5 (ablation) | TRAM · ICM->O | AUC | No clear effect |
| give BUDoPT five target bonafide samples in training | BUDoPT · Protocol 2 (avg) | HTER (%): 0.91 -> 0.96 | No clear effect |
| use TRAM finetuning-free relation modeling instead of OTA | finetuning-free TTA · Protocol 2 (avg) | HTER (%): 3.21 -> 0.81 | Improved |
Closest papers by meaning
- Multi-Domain Learning for Updating Face Anti-Spoofing Models (ECCV 2022 · similarity 0.90)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.90)
- Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing (ECCV 2024 · similarity 0.90)
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.90)
- CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing (CVPR 2024 · similarity 0.88)
- Optimal Transport-Guided Source-Free Adaptation for Face Anti-Spoofing (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
