Research map / Label-efficient and robust learning / Recognition and long-tail
MDFG: Multi-Dimensional Fine-Grained Modeling for Fatigue Detection
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 MDFG, which defines multi-dimensional fine-grained fatigue labels (time, type, level) and automatically synthesizes intermediate fatigue-state Gaussian base classes from SmallLoss-filtered clustered GMMs, plus meta-learning, for eye-movement-based fatigue detection.
Problem
Fatigue detection relies on coarse-grained labels spanning minutes to hours and treats alert and fatigue as two separate distributions, missing short-term intermediate fatigue states and fatigue type/level diversity.
Approach
Cluster trustworthy alert and fatigue samples into subtypes, fit GMM base classes, and synthesize intermediate-state base classes via Gaussian linear combination weighted by a fatigue score; then quantify each segment's fatigue score by nearest-base-class matching and train a DNN with a pairwise meta-learning module.
Key results
- On reRLDD, MDFG reaches Acc 0.719 and Av acc 0.688, versus 0.631 Acc for TimesNet (best baseline F1/Acc differences reported as 8.8% gain).
- On reDROZY, MDFG reaches Acc 0.781 and Av acc 0.746, improving over Lee et al. (eye) by about 2.1% Acc and 4.9% Av acc.
- Compared with Mehmood et al. (10fps), MDFG improves Av acc by 12.2% and Av F1 by 13% on reRLDD (Table 6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add meta-learning (ablation) | MDFG · reRLDD | Av acc: 0.655 -> 0.688 | Helped |
| add meta-learning (ablation) | MDFG · reDROZY | Av acc: 0.730 -> 0.746 | Helped |
| use β groups [0,1] (2 groups) (ablation) | MDFG · reRLDD and reDROZY | Av acc: 0.608; 0.659 | Mixed |
| use β groups [0,0.5,1] (3 groups) (ablation) | MDFG · reRLDD and reDROZY | Av acc: 0.644; 0.723 | Mixed |
| use β groups [0,0.3,0.5,0.7,1] (5 groups) (ablation) | MDFG · reRLDD and reDROZY | Av acc: 0.611; 0.736 | Mixed |
| use β groups [0,0.1,0.3,0.5,0.7,0.9,1] (7 groups) (ablation) | MDFG · reRLDD and reDROZY | Av acc: 0.655; 0.730 | Helped |
| use 11 β groups from 0 to 1 by 0.1 (ablation) | MDFG · reRLDD and reDROZY | Av acc: 0.617; 0.691 | Mixed |
| replace pairwise meta-learning with MAML (ablation) | MDFG · reRLDD | Av acc: 0.637 vs 0.688 | Hurt |
| replace pairwise meta-learning with Meta Networks (ablation) | MDFG · reRLDD | Av acc: 0.664 vs 0.688 | Hurt |
| replace pairwise meta-learning with Relation Networks (ablation) | MDFG · reRLDD | Av acc: 0.659 vs 0.688 | Hurt |
Surprising findings
- Setting fatigue-score groups to 0.4 and 0.6 (close to 0.5) degraded performance, as near-boundary intermediate bases cause misclassification.
- Adding the 0.1 and 0.9 intermediate bases to the 0.5 setting improved results, but the Table 2 grouping with 5 groups scored lower than with 3 or 7 groups.
Limitations
- Relies on indirect validation since no fine-grained ground-truth labels exist for intermediate states.
- Performance is sensitive to the trustworthy threshold and fatigue-score group settings near 0.5.
- Uses only eye-movement features; multimodal integration is left for future work.
Experimental setup
Datasets: reRLDD reDROZY
Benchmarks: reRLDD reDROZY
Metrics: Acc F1 Av acc Av F1
Baselines: Eye4feature Eye16feature HMLSTM FacialUnits Lee et al. Khandare et al. Mehmood et al. iTransformer TimesNet TimesURL MILLET CNN-LSTM
Builds on: HMLSTM SmallLoss Relation Networks
Closest papers by meaning
- BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation (AAAI 2025 · similarity 0.83)
- Task-Specific Gradient Adaptation for Few-Shot One-Class Classification (CVPR 2025 · similarity 0.80)
- Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models (NeurIPS 2026 · similarity 0.78)
- Aeolus: A Multi-structural Flight Delay Dataset (NeurIPS 2025 · similarity 0.78)
- Dynamic Pseudo Labeling via Gradient Cutting for High-Low Entropy Exploration (CVPR 2025 · similarity 0.78)
- Accident Anticipation via Temporal Occurrence Prediction (NeurIPS 2025 · similarity 0.78)
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.