Research map / Label-efficient and robust learning / Recognition and long-tail
AHAN: Asymmetric Hierarchical Attention Network for Identical Twin Face Verification
Hoang-Nhat Nguyen
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 Hierarchical Cross-Attention (HCA) (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 81.2 -> 87.1 | Improved |
| add Facial Asymmetry Attention Module (FAAM) (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 81.2 -> 84.7 | Improved |
| add TA-PWCA twin-distractor regularization (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 81.2 -> 89.9 | Improved |
| add FAAM on top of HCA (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 87.1 -> 90.4 | Improved |
| add TA-PWCA on top of HCA + FAAM (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 90.4 -> 92.3 | Improved |
| add HCA on top of FAAM + TA-PWCA (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 90.7 -> 92.3 | Improved |
| add FAAM on top of HCA + TA-PWCA (ablation) | AHAN (ViT-B baseline) · ND TWIN (Table 4 ablation, labeled hard twin) | TVA (%): 91.8 -> 92.3 | Improved |
| replace ArcFace (ResNet-100) with AHAN | face verification · ND TWIN (twin verification) | accuracy (%): 88.9 -> 92.3 | Improved |
| replace ArcFace with AHAN on hard twin pairs | face verification · ND TWIN (hard twin verification) | accuracy (%): 85.3 -> 88.5 | Improved |
| replace ArcFace with AHAN on general verification | face verification · ND TWIN (general verification) | accuracy (%): 98.5 -> 99.1 | Improved |
| use plain ViT transformer face model (TransFace) instead of CNN ArcFace | baseline face recognition · ND TWIN (twin verification) | accuracy (%): 88.9 -> 85.2 | Hurt |
| use twins as distractors instead of random pairs, hard negatives or siblings in regularization (ablation) | AHAN regularization · ND TWIN (hard twin verification) | accuracy (%) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 5 of 5 shared results
- MagFace — MagFace: A Universal Representation for Face Recognition and Quality Assessment (CVPR 2021) · 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
- BioMetricNet: deep unconstrained face verification through learning of metrics regularized onto Gaussian distributions (ECCV 2020 · similarity 0.85)
- Attentional Feature-Pair Relation Networks for Accurate Face Recognition (ICCV 2019 · similarity 0.85)
- DVG-Face: Dual Variational Generation for Heterogeneous Face Recognition (TPAMI 2021 · similarity 0.85)
- FM2u-Net: Face Morphological Multi-Branch Network for Makeup-Invariant Face Verification (CVPR 2020 · similarity 0.85)
- SapiensID: Foundation for Human Recognition (CVPR 2025 · similarity 0.83)
- QCS:Feature Refining from Quadruplet Cross Similarity for Facial Expression Recognition (AAAI 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.
