Research map / Label-efficient and robust learning / Recognition and long-tail
Dynamic Stereotype Theory Induced Micro-expression Recognition with Oriented Deformation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SODA4MER, an apex-free micro-expression recognition model that pretrains a self-supervised oriented deformation estimator on VoxCeleb, adds muscle-group-prior gated temporal variance Gaussian features, detects the apex via triplet contrastive loss, and models activation and decay phases with Bi-LSTMs guided by dynamic stereotype theory.
Problem
Micro-expression recognition suffers from scarce samples, difficulty modeling nearly imperceptible facial movements, and reliance on apex frame annotations that are often unavailable or lost under video resampling.
Approach
Replaces onset-apex optical flow with self-supervised local deformation estimates (keypoints and Jacobians) refined by a gated temporal-variance Gaussian over facial landmark muscle-group priors, detects the apex contrastively as the frame of maximal local deformation, and encodes activation and decay phases separately following dynamic stereotype theory.
Key results
- On SMIC-HS three-class, SODA4MER reaches UF1 0.8855 and UAR 0.8881, beating the second-best method by 1.4% UF1 and 2.4% UAR.
- On CASME II three-class, SODA4MER gets UF1 0.8870, below MERASTC (0.9330), so it is not best on that dataset.
- Ablation on SMIC-HS: removing GTVG drops UF1 from 0.8855 to 0.6105 range effects; removing DPTM drops UF1 to 0.4140 and UAR to 0.4327.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SODA4MER instead of the strongest baseline, Shallow I3D | SODA4MER · CASME II, three-class | UF1: 0.9370 -> 0.8870 | Hurt |
| Use SODA4MER instead of the strongest baseline, MERASTC | SODA4MER · CASME II, three-class | UAR: 0.950 -> 0.8809 | Hurt |
| Use SODA4MER instead of MiNet | SODA4MER · SMIC-HS, three-class | UF1: 0.8730 -> 0.8855 | Helped |
| Use SODA4MER instead of MiNet | SODA4MER · SMIC-HS, three-class | UAR: 0.8670 -> 0.8881 | Helped |
| Use SODA4MER instead of MMNet | SODA4MER · CASME II, five-class | UF1: 0.8676 -> 0.8141 | Hurt |
| Use SODA4MER instead of MMNet | SODA4MER · CASME II, five-class | ACC: 0.8835 -> 0.8418 | Hurt |
| Use SODA4MER instead of CMNet | SODA4MER · SAMM, five-class | UF1: 0.7719 -> 0.7893 | Helped |
| Use SODA4MER instead of STRCN-G | SODA4MER · SAMM, five-class | ACC: 0.7860 -> 0.8030 | Helped |
| Add GTVG (ablation) | SODA4MER · SMIC-HS, three-class | UF1: 0.6105 -> 0.8855 | Helped |
| Add GTVG (ablation) | SODA4MER · SMIC-HS, three-class | UAR: 0.6360 -> 0.8881 | Helped |
| Add DPTM (ablation) | SODA4MER · SMIC-HS, three-class | UF1: 0.4140 -> 0.8855 | Helped |
| Add DPTM (ablation) | SODA4MER · SMIC-HS, three-class | UAR: 0.4327 -> 0.8881 | Helped |
Surprising findings
- Forcing the ground-truth apex as the detected apex degrades performance, which the authors attribute to disruption of the temporal pattern.
- Adding the large-scale CASME3 dataset to training worsened performance due to sample distribution differences.
Limitations
- Sensitive to certain hyperparameters, per the authors.
- Residual ME-unrelated deformation noise and occasional keypoint misalignment in qualitative failure cases.
- Model has 126.1M parameters, which the authors describe as complex.
Experimental setup
Datasets: VoxCeleb CASME II SAMM SMIC-HS CASME3
Benchmarks: CASME II SMIC-HS SAMM
Metrics: UF1 UAR ACC
Baselines: MERASTC GLEFFN FRL-DGT MiNet Micron-BERT OFF-ApexNet STSTNet CMNet
Builds on: Monkey-style first-order motion model keypoint deformation Dynamic Stereotype Theory focal loss VGG perceptual loss
Closest papers by meaning
- Region-Aware Instance Consistency Learning for Micro-Expression Recognition (CVPR 2026 · similarity 0.82)
- TEASER: Token Enhanced Spatial Modeling for Expressions Reconstruction (ICLR 2025 · similarity 0.78)
- PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization (CVPR 2024 · similarity 0.77)
- Learning Adaptive Spatial Coherent Correlations for Speech-Preserving Facial Expression Manipulation (CVPR 2024 · similarity 0.77)
- Lifting Scheme-Based Implicit Disentanglement of Emotion-Related Facial Dynamics in the Wild (AAAI 2025 · similarity 0.77)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.77)
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.