Research map / Label-efficient and robust learning / Noisy labels
LA-Net: Landmark-Aware Learning for Reliable Facial Expression Recognition under Label Noise
Zhiyu Wu, Jinshi Cui
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 label distribution estimation using expression space only (ablation) | LA-Net baseline · RAF-DB (30% noise) | accuracy (%): 75.50 -> 79.17 | Improved |
| add landmark information to label distribution estimation (ablation) | LA-Net · RAF-DB (30% noise) | accuracy (%): 79.17 -> 82.37 | Improved |
| add landmark information to label distribution estimation (ablation) | LA-Net · RAF-DB (original) | accuracy (%): 88.43 -> 89.04 | Improved |
| add expression-landmark contrastive loss with one-hot labels (ablation) | LA-Net (LD+LM) · RAF-DB (original) | accuracy (%): 89.04 -> 90.48 | Improved |
| use pseudo-labels in EL Loss pair selection instead of one-hot labels (ablation) | LA-Net · RAF-DB (30% noise) | accuracy (%): 83.21 -> 85.33 | Improved |
| use pseudo-labels in EL Loss pair selection instead of one-hot labels (ablation) | LA-Net · RAF-DB (original) | accuracy (%): 90.48 -> 90.81 | Improved |
| vary number of neighbors K (K=12 vs 8) (ablation) | LA-Net · RAF-DB original vs 30% noise | accuracy (%) | Mixed |
| replace EAC with LA-Net under label noise | FER noise-tolerant models · AffectNet (10%/20%/30% noise) | accuracy (%): 58.91 -> 60.82 at 30% noise | Improved |
| apply LA-Net instead of baseline at 30% noise | ResNet-18 · AffectNet (30% noise) | accuracy (%): 52.16 -> 60.82 | Improved |
| train on mixed RAF-DB and AffectNet with inconsistent labels using LA-Net instead of LDLVA | FER cross-dataset · AffectNet 7-class | accuracy (%): 62.89 -> 65.43 | Improved |
| LA-Net label distributions instead of one-hot labels as descriptor | FER annotation · AffectNet user study | JS divergence to human perception: 0.2030 -> 0.0909 | Improved |
| replace TransFER with LA-Net (ResNet-18) on RAF-DB original | FER · RAF-DB (original) | accuracy (%): 90.91 vs 90.81 | No clear effect |
Later papers that compared with this method
3 later accepted papers report LA-Net as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Noise-Robust Facial Expression Recognition via Mamba-driven Neighbor Weight Refinement (ECCV 2026 · better on 6 of 6 shared results)
- Dynamic Label Noise Suppression with Optimal Teacher Pool for Facial Expression Recognition (CVPR 2026 · better on 4 of 5 shared results)
- Norface: Improving Facial Expression Analysis by Identity Normalization (ECCV 2024 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CLEX: Complementary Label Exchange Learning for Noisy Facial Expression Recognition (CVPR 2026 · similarity 0.88)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.88)
- Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning (TPAMI 2021 · similarity 0.88)
- Noise-Robust Facial Expression Recognition via Mamba-driven Neighbor Weight Refinement (ECCV 2026 · similarity 0.87)
- LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · similarity 0.86)
- Open-Set Facial Expression Recognition (AAAI 2024 · 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.
