Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding and Mitigating Annotation Bias in Facial Expression Recognition
Yunliang Chen, Jungseock Joo
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 AU-calibrated triplet loss (AUC-FER) | ResNet-50 baseline · ExpW to CFD happy | CV discrimination score: 0.059 ± 0.035 -> 0.006 ± 0.020 | Improved |
| add AU-calibrated triplet loss (AUC-FER) | MobileNetV2 baseline · ExpW to CFD happy (8,000) | CV discrimination score: 0.079 ± 0.009 -> 0.028 ± 0.029 | Improved |
| add domain independent training | ResNet-50 baseline · ExpW to CFD happy | CV discrimination score: 0.059 ± 0.035 -> 0.029 ± 0.015 | Improved |
| add gradient projection | ResNet-50 baseline · ExpW to CFD happy | CV discrimination score: 0.059 ± 0.035 -> 0.036 ± 0.014 | Improved |
| add domain discriminative training | ResNet-50 baseline · ExpW to CFD happy | CV discrimination score: 0.059 ± 0.035 -> 0.076 ± 0.024 | Hurt |
| add uniform confusion | MobileNetV2 baseline · ExpW to CFD happy (8,000) | CV discrimination score: 0.079 ± 0.009 -> 0.085 ± 0.021 | Hurt |
| add gradient projection | MobileNetV2 baseline · ExpW to CFD happy (8,000) | CV discrimination score: 0.079 ± 0.009 -> 0.070 ± 0.036 | No clear effect |
| relabel training data so happy proportion is equal across genders given AUs (ablation) | ResNet-50 trained on ExpW · ExpW to CFD happy | CV discrimination score: 0.0574 -> 0.0052 | Improved |
| train on web-annotated rather than lab-controlled data (annotation bias) (ablation) | happy labels · AU6=1 gender gap | female minus male happy proportion: ExpW 0.124, RAF-DB 0.153 vs CFD 0.016, KDEF 0.072 | Hurt |
| use automatic labels instead of manual labels (ablation) | AffectNet · AU6=1 gender gap | female minus male happy proportion: 0.145 -> 0.086 | No clear effect |
Closest papers by meaning
- Facial Expression Recognition with Inconsistently Annotated Datasets (ECCV 2018 · similarity 0.90)
- EmotioNet: An Accurate, Real-Time Algorithm for the Automatic Annotation of a Million Facial Expressions in the Wild (CVPR 2016 · similarity 0.89)
- FAIR-FER: A Latent Alignment Approach for Mitigating Bias in Facial Expression Recognition (Student Abstract) (AAAI 2024 · similarity 0.88)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.88)
- Open-Set Facial Expression Recognition (AAAI 2024 · similarity 0.86)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · 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.
