Research map / Image generation / Editing and style transfer
Goldilocks Test Sets for Face Verification
Haiyu Wu, Sicong Tian, Aman Bhatta, Jacob Gutierrez, Grace Bezold, Genesis Argueta, Karl Ricanek, Michael C. King, Kevin Bowyer
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 random MORPH5 pairs instead of attribute-targeted pairs (ablation) | Hadrian/Eclipse · Random MORPH5 test set | accuracy: 99.82 vs 91.72 (Hadrian) and 81.73 (Eclipse), ArcFace MS1MV2 | Hurt |
| assign genuine pairs of same identity to different folds (overlap) instead of disjoint (ablation) | Hadrian · Hadrian | accuracy: 91.72 -> 91.72 (ArcFace), 91.48 -> 91.45 (CurricularFace) | No clear effect |
| identity overlap across folds instead of disjoint (ablation) | Eclipse · Eclipse | accuracy: 81.73 -> 81.68 (ArcFace), 82.67 -> 82.60 (MagFace) | No clear effect |
| identity overlap across folds instead of disjoint (ablation) | ND-Twins · ND-Twins | accuracy: 68.55 -> 68.70 (ArcFace), 65.73 -> 65.82 (UniFace) | No clear effect |
| compare African-American male vs Caucasian male subsets | ArcFace · Hadrian | accuracy: 97.17 (AAM) vs 86.27 (CM) | Mixed |
| compare Caucasian female vs African-American male subsets | ArcFace · Eclipse | accuracy: 95.00 (AAM) vs 65.33 (CF) | Mixed |
| train on larger/cleaner Glint360K instead of MS1MV2 | ArcFace · ND-Twins | accuracy: 68.55 -> 84.67 | Improved |
| train on Glint360K instead of MS1MV2 | ArcFace · Hadrian | accuracy: 91.72 -> 95.68 | Improved |
| do not restrict image occurrence across pairs (ablation) | FR models · ND-Twins | accuracy: none better than 50% without constraint vs 71.56 average with it | Hurt |
| do not restrict image occurrence across pairs (ablation) | FR models · Hadrian/Eclipse | accuracy: less than 55% without constraint vs 92.62 and 82.81 average with it | Hurt |
Closest papers by meaning
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.86)
- OneFace: One Threshold for All (ECCV 2022 · similarity 0.85)
- UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition (NeurIPS 2023 · similarity 0.84)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.83)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.83)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.83)
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.
