Research map / Label-efficient and robust learning / Recognition and long-tail
Efficient Facial Feature Learning with Wide Ensemble-Based Convolutional Neural Networks
Henrique Siqueira, Sven Magg, Stefan Wermter
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 |
|---|---|---|---|
| use shared representations and branch at level 3 | ESR-4 on CK+ · CK+ | accuracy: 85.5 ± 3.5 % -> 89.4 ± 2.2 % | Improved |
| branch at level 4 instead of level 3 (ablation) | ESR-4 on CK+ · CK+ | accuracy: 89.4 ± 2.2 % -> 88.5 ± 3.8 % | Mixed |
| branch at level 4 instead of level 3 (ablation) | ESR-4 on CK+ · CK+ | trainable parameters: $355.104$ -> $243.936$ | Improved |
| use a traditional ensemble rather than shared representations | four-network ensemble on CK+ · CK+ | accuracy: 89.2 ± 1.2 % vs 89.4 ± 2.2 % | No clear effect |
| train ensemble branches interleaved rather than sequentially (ablation) | ESR on CK+ · CK+ | accuracy | Hurt |
| train new branches using shared layers as transferred feature knowledge | ESR on CK+ · CK+ | training convergence: twice as accurate; around epoch 50 | Improved |
| collectively classify with nine branches instead of using a single branch | ESR-9 · AffectNet discrete emotions | accuracy: 58.0\% -> 59.3\% | Improved |
| train ESR-9 on balanced subsets capped at 5000 samples per emotion (ablation) | ESR-9 · AffectNet discrete emotions | accuracy | Improved |
| fine-tune the AffectNet-trained ESR-9 on FER+ | ESR-9 · FER+ | accuracy: $57.92\%$ -> $87.153\%$ | Improved |
| lower the maximum training samples per emotion category (ablation) | ESR-9 · FER+ | overall accuracy | Hurt |
| raise the maximum training samples per emotion category too high (ablation) | ESR-9 · FER+ | ensemble diversity | Hurt |
| add arousal and valence output neurons while retaining the discrete-emotion network | ESR-9 · AffectNet continuous affect | RMSE: 0.33 and 0.36 RMSE | Improved |
Closest papers by meaning
- FERV39k: A Large-Scale Multi-Scene Dataset for Facial Expression Recognition in Videos (CVPR 2022 · similarity 0.87)
- Deep Structured Learning for Facial Action Unit Intensity Estimation (CVPR 2017 · similarity 0.87)
- Reliable Crowdsourcing and Deep Locality-Preserving Learning for Expression Recognition in the Wild (CVPR 2017 · similarity 0.86)
- Knowledge Augmented Deep Neural Networks for Joint Facial Expression and Action Unit Recognition (NeurIPS 2020 · similarity 0.86)
- SynFER: Towards Boosting Facial Expression Recognition with Synthetic Data (ICCV 2025 · similarity 0.86)
- A Compact Embedding for Facial Expression Similarity (CVPR 2019 · 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.
