Research map / Label-efficient and robust learning / Noisy labels
D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition.
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 D2SP, a plug-and-play dual-stage dynamic purification framework that prunes low-quality samples (coarse-grained pruning via learned sample weights) and corrects mislabeled samples (fine-grained correction via prediction stability) for dynamic facial expression recognition.
Problem
Dynamic facial expression recognition datasets contain noisy samples of two kinds: low-quality data (occlusion, dim lighting, blur) and mislabeled data from annotation bias, which existing noisy-label methods treat as a single noise type.
Approach
Separates the two noise types and handles them sequentially: a learned per-sample weight module with threshold pruning removes low-quality samples (CGP), then label-prediction and score stability across recent epochs decide which samples to relabel (FGC), with the purification recomputed every epoch.
Key results
- On FERV39k, D2SP improved WAR/UAR by an average of 3.97%/2.45% across six DFER methods, with ResNet50_LSTM gaining 7.82% WAR (40.75 to 48.57).
- On DFEW, average gains were 1.67% WAR and 0.84% UAR; on MAFW, 1.20% WAR and 0.56% UAR.
- Ablation on FERV39k with ResNet18_LSTM: CGP alone +3.54% WAR, FGC alone +2.51% WAR, combined +4.73% WAR (42.59 to 47.32).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add D2SP to ViT_LSTM | ViT_LSTM · FERV39k | WAR: 48.67 -> 50.5 | Helped |
| Add D2SP to ViT_LSTM | ViT_LSTM · DFEW | WAR: 67.07 -> 69.2 | Helped |
| Add D2SP to ViT_LSTM | ViT_LSTM · MAFW | WAR: 41.6 -> 42.58 | Helped |
| Add CGP (coarse-grained pruning) (ablation) | ResNet18_LSTM · FERV39k | WAR: 42.59 -> 46.13 | Helped |
| Add CGP (coarse-grained pruning) (ablation) | ResNet18_LSTM · DFEW | WAR: 63.85 -> 64.25 | Helped |
| Add FGC (fine-grained correction) (ablation) | ResNet18_LSTM · FERV39k | WAR: 42.59 -> 45.1 | Helped |
| Add FGC (fine-grained correction) (ablation) | ResNet18_LSTM · DFEW | WAR: 63.85 -> 64.42 | Helped |
| Use CGP and FGC together with dynamic purification (ablation) | ResNet18_LSTM · FERV39k | WAR: 42.59 -> 47.32 | Helped |
| Use CGP and FGC together with dynamic purification (ablation) | ResNet18_LSTM · DFEW | WAR: 63.85 -> 64.72 | Helped |
| Disable dynamic purification while retaining CGP and FGC (ablation) | ResNet18_LSTM · FERV39k | WAR: 42.59 -> 41.51 | Hurt |
| Disable dynamic purification while retaining CGP and FGC (ablation) | ResNet18_LSTM · DFEW | UAR: 51.32 -> 48.73 | Hurt |
| Vary CGP threshold λ from 0.5 to 0.9 (ablation) | ResNet18_LSTM + D2SP · FERV39k and DFEW | WAR: FERV39k: 46.85, 46.9, 47.32, 46.32, 45.75; DFEW: 64.29, 64.72, 64.03, 62.37, 61.77 | Mixed |
Surprising findings
- Applying only the FGC stage or only the CGP stage gives weaker results, and on DFEW the single-stage variants slightly lower UAR, so the two stages are only beneficial jointly.
- Static purification (using one round's purified set) performed worse than dynamic per-epoch purification on both DFEW and FERV39k.
Limitations
- Hyperparameters (lambda, tau, t) are tuned per dataset and per model, with optimal values differing across datasets.
- Gains on DFEW and MAFW are modest (often under 1-2% WAR), and the paper reports no variance or significance testing.
- The paper provides only a qualitative analysis of why low-weight samples are low quality, with limited theoretical justification.
Experimental setup
Datasets: DFEW FERV39k MAFW
Benchmarks: DFEW FERV39k MAFW
Metrics: WAR UAR
Baselines: Res18_LSTM Res50_LSTM VGG13_LSTM VGG16_LSTM Former-DFER ViT_LSTM
Builds on: MentorNet DivideMix Former-DFER
Closest papers by meaning
- D^3FER: Dual Channel and Dual Branch Network for Robust Facial Expression Recognition under Dual Challenges (CVPR 2026 · similarity 0.90)
- OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem (AAAI 2025 · similarity 0.87)
- Lifting Scheme-Based Implicit Disentanglement of Emotion-Related Facial Dynamics in the Wild (AAAI 2025 · similarity 0.86)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.85)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.83)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.83)
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.