Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Multi-Task Learning to Recognise Subtle Facial Expressions of Mental States
Guosheng Hu, Li Liu, Yang Yuan, Zehao Yu, Yang Hua, Zhihong Zhang, Fumin Shen, Ling Shao, Timothy Hospedales, Neil Robertson, Yongxin Yang
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 trace-norm soft sharing (LAF) with landmark detection task instead of single-task training | ResNet-34 · LSEMSW mental state recognition | accuracy (%): 28.39 -> 33.43 | Improved |
| add adversarial domain alignment (ADA) to LAF soft sharing (ablation) | ResNet-34+LAF · LSEMSW mental state recognition | accuracy (%): 33.43 -> 36.72 | Improved |
| add adversarial domain alignment (ADA) to LAF soft sharing (ablation) | ResNet-34+LAF · 300-W full set | normalised mean error (%): 4.67 -> 4.64 | Improved |
| use Tucker or TT trace norm instead of LAF (ablation) | ResNet-34 soft sharing · LSEMSW mental state recognition | accuracy (%): 33.43 vs 33.39 vs 33.41 | No clear effect |
| replace hard sharing (first 6 layers shared) with automatic soft sharing LAF | ResNet-34 MTL · LSEMSW mental state recognition | accuracy (%): 30.07 -> 33.43 | Improved |
| share more layers in hard sharing (6 to 32 layers) (ablation) | ResNet-34 hard sharing · LSEMSW mental state recognition | accuracy (%): 30.07 -> 24.69 | Hurt |
| replace cross stitch MTL with trace-norm LAF sharing | ResNet-34 MTL · LSEMSW mental state recognition | accuracy (%): 30.96 -> 33.43 | Improved |
| add ADA to hard sharing RN(6) (ablation) | ResNet-34 hard sharing · LSEMSW mental state recognition | accuracy (%): 30.07 -> 33.97 | Improved |
| use l1 loss instead of l2 loss for landmark regression (ablation) | RN+LAF+ADA · 300-W full set | normalised mean error (%): 4.76 -> 4.64 | Improved |
| fine-tune from LSEMSW pretraining instead of training from scratch (ablation) | RN+LAF+ADA · Oulu-Casia NIR&Vis | accuracy (%): 76.0 -> 87.1 | Improved |
| fine-tune from LSEMSW pretraining instead of training from scratch (ablation) | RN+LAF+ADA · CK+ | accuracy (%): 86.3 -> 96.4 | Improved |
| add LAF and ADA to fine-tuned ResNet (ablation) | RN (FT) · Oulu-Casia NIR&Vis | accuracy (%): 82.9 -> 87.1 | Improved |
Closest papers by meaning
- A Compact Embedding for Facial Expression Similarity (CVPR 2019 · similarity 0.88)
- EmotioNet: An Accurate, Real-Time Algorithm for the Automatic Annotation of a Million Facial Expressions in the Wild (CVPR 2016 · similarity 0.88)
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.88)
- When Facial Expression Recognition Meets Few-Shot Learning: A Joint and Alternate Learning Framework (AAAI 2022 · similarity 0.88)
- Multi-Label Compound Expression Recognition: C-EXPR Database & Network (CVPR 2023 · similarity 0.87)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · 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.
