Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Face Recognition by Clustering Unlabeled Faces in the Wild
Aruni RoyChowdhury, Xiang Yu, Kihyuk Sohn, Erik Learned-Miller, Manmohan Chandraker
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 GCN-clustered pseudo-labeled unlabeled faces | CosFace face recognition retrained with pseudo-labels · CFP-fp | accuracy: 92.37 -> 95.51 | Improved |
| add GCN-clustered pseudo-labeled unlabeled faces | CosFace face recognition retrained with pseudo-labels · IJB-A verification | TAR@FAR=1e-3: 80.23 -> 87.60 | Improved |
| add GCN-clustered pseudo-labeled unlabeled faces | CosFace face recognition retrained with pseudo-labels · IJB-A identification | Rank-1: 92.66 -> 94.11 | Improved |
| add uncertainty-weighted loss with gamma=1 on pseudo-labels (ablation) | CosFace face recognition retrained with pseudo-labels · IJB-A verification | TAR@FAR=1e-3: 87.60 -> 87.97 | Improved |
| vary gamma in uncertainty-weighted loss (1/3 to 3) (ablation) | CosFace face recognition retrained with pseudo-labels · IJB-A, CFP-fp | several: 87.93 to 88.43 TAR@FAR=1e-3 for gamma 1/3 to 3 | No clear effect |
| use gamma=1/2 instead of other gamma values (ablation) | CosFace face recognition retrained with pseudo-labels · IJB-A verification | TAR@FAR=1e-3: 84.13 vs. 87.97 (gamma=1) | Hurt |
| use Weibull 95% instead of naive Otsu threshold for overlapping identity separation (ablation) | Max-logit OOD detection · Overlapping identity separation | false positive rate: 6.2% -> 2.33% | Improved |
| use Weibull instead of Gaussian model for overlap separation (ablation) | Max-logit OOD detection · Overlapping identity separation | false positive / negative rate: 2.01% vs. 2.33% FP, 0.51% vs. 0.50% FN | No clear effect |
| use classification-margin instead of softmax entropy as uncertainty metric (ablation) | Noise model for clusters · Clustering error detection | average precision: 88.29% -> 95.16% | Improved |
Closest papers by meaning
- Learning to Cluster Faces on an Affinity Graph (CVPR 2019 · similarity 0.87)
- On Mitigating Hard Clusters for Face Clustering (ECCV 2022 · similarity 0.85)
- Clustering Millions of Faces by Identity (TPAMI 2017 · similarity 0.84)
- Invariant Information Clustering for Unsupervised Image Classification and Segmentation (ICCV 2019 · similarity 0.83)
- Learning to Cluster Faces via Confidence and Connectivity Estimation (CVPR 2020 · similarity 0.82)
- Face-Specific Data Augmentation for Unconstrained Face Recognition (IJCV 2019 · similarity 0.82)
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.
