Research map / Architectures and efficiency / Attention and transformers
AUFormer: Vision Transformers are Parameter-Efficient Facial Action Unit Detectors
Kaishen Yuan, Zitong Yu, Xin Liu, Weicheng Xie, Huanjing Yue, Jingyu Yang
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 |
|---|---|---|---|
| replace full fine-tuning with PETL (Convpass) (ablation) | ViT-B/16 AU detector · BP4D | average F1: 61.9 -> 63.6 | Improved |
| add collaboration mechanism over a single Convpass (ablation) | AUFormer · BP4D | average F1: 63.6 -> 64.3 | Improved |
| add MRF and CA operators in MoKE (ablation) | AUFormer · BP4D | average F1: 64.3 -> 65.1 | Improved |
| replace WA-Loss with MDWA-Loss (gamma and m) (ablation) | full fine-tuned ViT baseline · BP4D | average F1: 61.9 -> 63.4 | Improved |
| add truncation margin m in MDWA-Loss (ablation) | AUFormer · BP4D | average F1: 65.6 -> 66.2 | Improved |
| add difficulty-aware gamma in MDWA-Loss (ablation) | AUFormer · BP4D | average F1: 65.1 -> 65.6 | Improved |
| use Convpass instead of Adapter, LoRA or VPT as adaptation module (ablation) | PETL AU detector · BP4D | average F1: 63.1 / 63.0 / 62.6 -> 63.6 | Improved |
| use dilated rates 1, 3, 5 instead of rate 1 only (ablation) | MRF operator · BP4D | average F1: 65.6 -> 66.2 | Improved |
| add dilation rate 2 (1,2,3 vs 1,3) (ablation) | MRF operator · BP4D | average F1: 65.9 -> 65.9 | No clear effect |
| enlarge CA neighborhood from 3x3 to 5x5 or 7x7 (ablation) | CA operator · BP4D | average F1: 66.2 -> 65.9 (5x5), 66.0 (7x7) | Hurt |
| increase truncation margin m from 0.1 to 0.2 (ablation) | MDWA-Loss · BP4D | average F1: 66.2 -> 65.4 | Hurt |
| apply AUFormer in BP4D to DISFA transfer vs FG-Net | AUFormer · BP4D to DISFA | average F1: 54.4 (FG-Net) vs 43.6 | Hurt |
Closest papers by meaning
- Breaking Spurious Correlations: Uncertainty-Driven Causal Transformers for AU Detection (CVPR 2026 · similarity 0.90)
- Knowledge-Spreader: Learning Semi-Supervised Facial Action Dynamics by Consistifying Knowledge Granularity (ICCV 2023 · similarity 0.89)
- Knowledge Augmented Deep Neural Networks for Joint Facial Expression and Action Unit Recognition (NeurIPS 2020 · similarity 0.89)
- Facial Action Unit Intensity Estimation via Semantic Correspondence Learning with Dynamic Graph Convolution (AAAI 2020 · similarity 0.88)
- AUEditNet: Dual-Branch Facial Action Unit Intensity Manipulation with Implicit Disentanglement (CVPR 2024 · similarity 0.86)
- Trend-Aware Supervision: On Learning Invariance for Semi-supervised Facial Action Unit Intensity Estimation (AAAI 2024 · similarity 0.86)
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.
