Research map / Label-efficient and robust learning / Recognition and long-tail
TopoFR: A Closer Look at Topology Alignment on Face Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TopoFR, a face recognition model that aligns persistent-homology topological structure between input and latent spaces (PTSA with random perturbation) and mines hard samples via a structure damage score (SDE), reporting SOTA on face benchmarks.
Problem
Face recognition models trained on large-scale data lose input-space structure information in the latent space; directly aligning topological structure causes a structure collapse (overfitting) problem, and hard/low-quality samples further damage latent topology.
Approach
PTSA perturbs the latent space with random augmentations (RSP) and aligns the persistent-homology (H0) pairwise-distance values of the input and perturbed latent spaces (ISA), avoiding bottleneck/Wasserstein distances; SDE uses a Gaussian-uniform mixture over prediction entropy plus focal-style weighting to compute a structure damage score that re-weights hard samples.
Key results
- On IJB-C (TAR@FAR=1e-4) with MS1MV2, R50 TopoFR reaches 96.49 vs 92.52 for R50 ArcFace.
- With Glint360K, R200 TopoFR reaches 96.71 at 1e-5 and 97.84 MFR-All-related IJB-C scores, surpassing TransFace-L (96.29 at 1e-5).
- TopoFR ranked second place on the ICCV21 MFR-Ongoing academic track at submission time (with R200 backbone, WebFace42M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TopoFR instead of UniFace | R200 face recognition trained on WebFace42M · MFR-Ongoing Mask | verification accuracy (%): 92.43 -> 93.96 | Helped |
| add RSP perturbations to ArcFace without PTSA (ablation) | R50 ArcFace · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 92.52 -> 92.44 | Hurt |
| align input and latent structures directly without RSP (ablation) | R50 ArcFace · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 92.52 -> 93.26 | Helped |
| add full PTSA (ablation) | R50 ArcFace · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 92.52 -> 95.34 | Helped |
| add GUM uncertainty weighting to PTSA (ablation) | TopoFR-P · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 95.34 -> 96.23 | Helped |
| add focal-loss weighting to PTSA (ablation) | TopoFR-P · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 95.34 -> 95.40 | Helped |
| combine SDE with PTSA rather than use focal weighting alone (ablation) | TopoFR · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 95.40 -> 96.49 | Helped |
| use SDE instead of earlier hard-sample mining strategies (ablation) | R100 ArcFace · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 96.18 -> 96.49 | Helped |
| replace ISA distance computation with Bottleneck distance (ablation) | R100 TopoFR · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 96.95 -> 96.76 | Hurt |
| replace ISA distance computation with 1-Wasserstein distance (ablation) | R100 TopoFR · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 96.95 -> 96.81 | Hurt |
| increase batch size from 128 to 256 (ablation) | R100 TopoFR · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 96.95 -> 96.91 | No clear effect |
| increase batch size from 128 to 512 (ablation) | R100 TopoFR · IJB-C (1e-4), MS1MV2 | verification accuracy (%): 96.95 -> 96.93 | No clear effect |
Surprising findings
- Directly aligning persistent-homology structure (TopoFR-A) makes the structure discrepancy drop to zero early in training, indicating a structure collapse/overfitting effect.
- Random data augmentations hurt plain ArcFace (TopoFR-R drops IJB-C 1e-4 from 92.52 to 92.44), but help when combined with structure alignment.
Limitations
- Training time increases about 1.16x over ArcFace.
- Hard-sample weighting may over-emphasize mislabeled samples in noisy large-scale datasets.
- Mainly validated on face recognition benchmarks; gains on saturated benchmarks (LFW, CFP-FP, AgeDB-30) are small.
Experimental setup
Datasets: MS1MV2 Glint360K WebFace42M
Benchmarks: LFW AgeDB-30 CFP-FP CPLFW CALFW IJB-C IJB-B MFR-Ongoing
Metrics: verification accuracy TAR@FAR=1e-4 TAR@FAR=1e-5 TAR@FAR=1e-6
Baselines: ArcFace CosFace AdaFace MagFace TransFace Partial FC UniFace MV-Softmax CurricularFace EHSM
Builds on: ArcFace Focal loss Persistent homology CosFace
Closest papers by meaning
- Attributes Shape the Embedding Space of Face Recognition Models (ICML 2025 · similarity 0.85)
- GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · similarity 0.84)
- Topograph: An Efficient Graph-Based Framework for Strictly Topology Preserving Image Segmentation (ICLR 2025 · similarity 0.84)
- Towards Scalable Topological Regularizers (ICLR 2025 · similarity 0.84)
- MANTRA: The Manifold Triangulations Assemblage (ICLR 2025 · similarity 0.84)
- Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning (NeurIPS 2026 · similarity 0.84)
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.