Research map / Label-efficient and robust learning / Recognition and long-tail
Pre-training Strategies and Datasets for Facial Representation Learning
Adrian Bulat, Shiyang Cheng, Jing Yang, Andrew Garbett, Enrique Sanchez, Georgios Tzimiropoulos
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 SwAV instead of MoCov2 for unsupervised pre-training (ablation) | face ResNet-50 · 300W (2% data) | NME (%): 11.9 -> 4.97 | Improved |
| use SwAV instead of MoCov2 for unsupervised pre-training (ablation) | face ResNet-50 · DISFA (2% data) | ICC: .280 -> .560 | Improved |
| use SwAV batch size 256 instead of 4096 (ablation) | SwAV pre-training · 300W (2% data) | NME (%): 4.97 vs 5.00 | No clear effect |
| pre-train on full VGG-Face instead of VGG-Face-small (ablation) | unsupervised SwAV · 300W (2% data) | NME: 5.46 -> 4.97 | Improved |
| pre-train on Large-Scale-Face (5M) instead of VGG-Face (ablation) | unsupervised SwAV · 300W (2% data) | NME: 4.97 vs 4.98 | No clear effect |
| pre-train on uncurated Flickr-Face instead of curated VGG-Face (ablation) | unsupervised SwAV · 300W (100% data) | NME: 3.85 vs 3.86 | No clear effect |
| unsupervised SwAV on faces instead of on ImageNet (ablation) | SwAV pre-training · 300W (2% data) | NME: 5.1 -> 4.97 | Improved |
| unsupervised face pre-training instead of training from scratch | ResNet-50 · 300W (2% data) | NME: 13.52 -> 4.97 | Improved |
| unsupervised face pre-training instead of supervised ImageNet pre-training | ResNet-50 · AFLW2000-3D (2% data) | NME: 6.15 -> 4.31 | Improved |
| unsupervised face pre-training instead of supervised VGG-F pre-training | ResNet-50 · IJB-B TAR@1e-4 (2% data) | TAR: 0.694 -> 0.744 | Improved |
| fine-tune the whole network instead of linear adaptation (ablation) | Ours (VGG-F) · DISFA (100% data) | ICC: .342 -> .598 | Improved |
| unsupervised pre-trained student for self-distillation instead of supervised/scratch (ablation) | self-distillation · 300W semi-supervised | NME | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 3FabRec — 3FabRec: Fast Few-Shot Face Alignment by Reconstruction (CVPR 2020) · better on 1 of 1 shared result
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
- Self-Supervised Facial Representation Learning with Facial Region Awareness (CVPR 2024 · similarity 0.90)
- General Facial Representation Learning in a Visual-Linguistic Manner (CVPR 2022 · similarity 0.90)
- Towards Universal Representation Learning for Deep Face Recognition (CVPR 2020 · similarity 0.89)
- LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · similarity 0.89)
- PrefAce: Face-Centric Pretraining with Self-Structure Aware Distillation (AAAI 2024 · similarity 0.88)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.87)
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.
