Research map / Label-efficient and robust learning / Recognition and long-tail
Navigating Label Ambiguity for Facial Expression Recognition in the Wild
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 Navigating Label Ambiguity (NLA), combining ambiguity-based adaptive loss weighting (NAW) with JSD consistency regularization on flipped images to handle both label noise and class imbalance in facial expression recognition.
Problem
Facial expression recognition in the wild suffers from label ambiguity and noisy labels together with class imbalance, and prior methods typically address only one of these.
Approach
Estimate each sample's ambiguity from the correlation between ground-truth and nearest-negative prediction scores, and weight the cross-entropy loss with multivariate Gaussian kernels (true/false-prediction forms, with an epoch-based covariance schedule), plus a Jensen-Shannon consistency term between original and flipped images, so later training emphasizes minority-class ambiguous samples while
Key results
- On RAF-DB with ResNet-18, NLA achieves 89.93% overall and 83.87% mean accuracy, above MEK (89.77% / 82.44%).
- On AffectNet, NLA reaches 67.06% overall accuracy, versus 65.73% for MEK.
- Under 30% label noise on RAF-DB, NLA reaches 86.71% overall accuracy, versus 85.33% for LA-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CE with NAW-CE, without adding regularization (ablation) | NLA · RAF-DB | overall accuracy: 87.42 -> 88.17 | Helped |
| Replace CE with NAW-CE, without adding regularization (ablation) | NLA · RAF-DB | mean accuracy: 78.53 -> 81.40 | Helped |
| Add JSD regularization to CE (ablation) | NLA · RAF-DB | overall accuracy: 87.42 -> 87.78 | Helped |
| Add JSD regularization to CE (ablation) | NLA · RAF-DB | mean accuracy: 78.53 -> 81.01 | Helped |
| Add L1 regularization to NAW-CE (ablation) | NLA · RAF-DB | overall accuracy: 88.17 -> 89.18 | Helped |
| Add L1 regularization to NAW-CE (ablation) | NLA · RAF-DB | mean accuracy: 81.40 -> 82.87 | Helped |
| Add CAM regularization to NAW-CE (ablation) | NLA · RAF-DB | overall accuracy: 88.17 -> 89.15 | Helped |
| Add CAM regularization to NAW-CE (ablation) | NLA · RAF-DB | mean accuracy: 81.40 -> 83.02 | Helped |
| Add JSD consistency regularization to NAW-CE (ablation) | NLA · RAF-DB | overall accuracy: 88.17 -> 89.93 | Helped |
| Add JSD consistency regularization to NAW-CE (ablation) | NLA · RAF-DB | mean accuracy: 81.40 -> 83.87 | Helped |
| Use NLA rather than MEK | ResNet-18 FER on RAF-DB · RAF-DB | overall accuracy: 89.77 -> 89.93 | Helped |
| Use NLA rather than MEK | ResNet-18 FER on AffectNet · AffectNet | overall accuracy: 65.73 -> 67.06 | Helped |
Limitations
- Covariance and mean vectors for the Gaussian kernels are set experimentally rather than learned.
- Evaluated only on FER datasets with a single ResNet-18 backbone.
Experimental setup
Datasets: RAF-DB AffectNet FERPlus MS-Celeb-1M
Benchmarks: RAF-DB AffectNet FERPlus
Metrics: overall accuracy mean accuracy
Baselines: CB loss BBN RUL EAC MEK LA-Net SCN Co-teaching
Builds on: JoCoR Class Activation Maps Balanced Softmax Decoupling
Closest papers by meaning
- CLEX: Complementary Label Exchange Learning for Noisy Facial Expression Recognition (CVPR 2026 · similarity 0.91)
- Open-Set Facial Expression Recognition (AAAI 2024 · similarity 0.88)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.87)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.86)
- FAIR-FER: A Latent Alignment Approach for Mitigating Bias in Facial Expression Recognition (Student Abstract) (AAAI 2024 · similarity 0.85)
- D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition. (CVPR 2025 · similarity 0.85)
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.