Research map / Video / Action recognition
Knowledge-Spreader: Learning Semi-Supervised Facial Action Dynamics by Consistifying Knowledge Granularity
Xiaotian Li, Xiang Zhang, Taoyue Wang, Lijun Yin
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 Transformer relation encoders with vanilla MLP (ablation) | Knowledge-Spreader · BP4D (50% labels) | F1: 64.5 -> 62.8 | Hurt |
| replace Transformer relation encoders with vanilla MLP (ablation) | Knowledge-Spreader · DISFA (50% labels) | F1: 61.6 -> 60.2 | Hurt |
| add knowledge distillation losses Ls and Lt (knowledge spreading) (ablation) | Knowledge-Spreader · BP4D and DISFA (50% labels) | F1: -2.0% on BP4D, -1.5% on DISFA when removed | Improved |
| add spatial knowledge distillation (SKD) (ablation) | Knowledge-Spreader · BP4D and DISFA (50% labels) | F1: -1.3% on BP4D, -1.1% on DISFA when removed | Improved |
| add temporal knowledge distillation (TKD) (ablation) | Knowledge-Spreader · BP4D and DISFA (50% labels) | F1: -0.9% on BP4D, -1.2% on DISFA when removed | Improved |
| add Perturbation-aware Pseudo-labeling module (ablation) | Knowledge-Spreader · BP4D and DISFA (50% labels) | F1: -0.8% on BP4D, -1.0% on DISFA when removed | Improved |
| replace self-supervised shuffle classifier with hard confidence threshold (ablation) | Knowledge-Spreader · BP4D and DISFA (50% labels) | F1: -0.3% on BP4D, -0.4% on DISFA | Hurt |
| filter pseudo labels with PPL instead of naive pseudo-labeling (ablation) | Knowledge-Spreader · BP4D (10% labels) | pseudo-label accuracy (%): 73.36% -> 76.35% | Improved |
| use KS with only 2% labels on BP4D versus supervised methods with 100% labels | Knowledge-Spreader · BP4D | F1: KS 10% labels 64.4 vs FAUDT 100% 64.2 | Improved |
| use KS instead of Pseudo-label, FixMatch or TCL under extremely few labels | semi-supervised AU detection · BP4D (1% labels) | F1: 55.6 -> 59.9 vs TCL | Improved |
Closest papers by meaning
- AUFormer: Vision Transformers are Parameter-Efficient Facial Action Unit Detectors (ECCV 2024 · similarity 0.89)
- Learning Facial Action Units From Web Images With Scalable Weakly Supervised Clustering (CVPR 2018 · similarity 0.89)
- ReactioNet: Learning High-Order Facial Behavior from Universal Stimulus-Reaction by Dyadic Relation Reasoning (ICCV 2023 · similarity 0.89)
- Trend-Aware Supervision: On Learning Invariance for Semi-supervised Facial Action Unit Intensity Estimation (AAAI 2024 · similarity 0.87)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.86)
- Learning Adaptive Spatial Coherent Correlations for Speech-Preserving Facial Expression Manipulation (CVPR 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.
