Research map / Label-efficient and robust learning / Recognition and long-tail
Margin-based face recognition losses
Shapes face embeddings with angular margins, adaptive margins and quality-aware losses.
124 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “face · person identification · facial”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
18 papers in 2023–24 and 9 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 1.4% to 0.7%.
Open questions the data points to
- CPLFW (verification accuracy (%)): best went from 93.48 (up to 2024) to 94.45 in 2025 — GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · 0 citations)
No longer reported by papers here: LFW (accuracy, last in 2022); CFP-FP (accuracy, last in 2022).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · DeepFashion: Powering Robust Clothes Recognition and Retrieval With Rich Annotations (CVPR 2016 · 1,931 citations)
- 2017 · SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR 2017 · 3,010 citations)
- 2018 · CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · 2,951 citations)
- 2019 · ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · 7,708 citations)
- 2020 · Circle Loss: A Unified Perspective of Pair Similarity Optimization (CVPR 2020 · 956 citations)
- 2021 · Frequency-Aware Discriminative Feature Learning Supervised by Single-Center Loss for Face Forgery Detection (CVPR 2021 · 324 citations)
- 2022 · Revisiting Skeleton-Based Action Recognition (CVPR 2022 · 825 citations)
- 2023 · Recognizability Embedding Enhancement for Very Low-Resolution Face Recognition and Quality Estimation (CVPR 2023 · 36 citations)
- 2024 · Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · 39 citations)
- 2025 · Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · 12 citations)
- 2026 · BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition (CVPR 2026 · 0 citations)
Earliest papers here
- DeepFashion: Powering Robust Clothes Recognition and Retrieval With Rich Annotations (CVPR 2016 · 1,931 citations)
- Compact Bilinear Pooling (CVPR 2016 · 896 citations)
- Ordinal Regression With Multiple Output CNN for Age Estimation (CVPR 2016 · 680 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
LFW · verification accuracy (%) (higher is better)
13 papers, 2018–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 99.46 | Virtual Class Enhanced Discriminative Embedding Learning (NeurIPS 2018 · 27 citations) |
| 2019 | 99.83 | ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · 7,708 citations) |
| 2020 | 99.85 | GroupFace: Learning Latent Groups and Constructing Group-Based Representations for Face Recognition (CVPR 2020 · 94 citations) |
LFW · accuracy (higher is better)
10 papers, 2016–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 99.5 | Latent Factor Guided Convolutional Neural Networks for Age-Invariant Face Recognition (CVPR 2016 · 163 citations) |
| 2018 | 99.73 | CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · 2,951 citations) |
| 2020 | 99.75 | Towards Universal Representation Learning for Deep Face Recognition (CVPR 2020 · 165 citations) |
| 2021 | 99.8 | Spherical Confidence Learning for Face Recognition (CVPR 2021 · 68 citations) |
| 2022 | 99.83 | CQA-Face: Contrastive Quality-Aware Attentions for Face Recognition (AAAI 2022 · 12 citations) |
CALFW · verification accuracy (%) (higher is better)
8 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 95.45 | ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · 7,708 citations) |
| 2020 | 96.2 | CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (CVPR 2020 · 590 citations) |
| 2025 | 96.94 | GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · 0 citations) |
CPLFW · verification accuracy (%) (higher is better)
8 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 92.08 | ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · 7,708 citations) |
| 2020 | 93.48 | More Information Supervised Probabilistic Deep Face Embedding Learning (ICML 2020 · 3 citations) |
| 2025 | 94.45 | GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · 0 citations) |
AgeDB · verification accuracy (%) (higher is better)
6 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 98.32 | CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (CVPR 2020 · 590 citations) |
| 2021 | 98.6 | Variational Prototype Learning for Deep Face Recognition (CVPR 2021 · 89 citations) |
| 2025 | 99.4 | GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · 0 citations) |
CFP-FP · accuracy (higher is better)
6 papers, 2020–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 98.64 | Towards Universal Representation Learning for Deep Face Recognition (CVPR 2020 · 165 citations) |
What has been tried
119 papers here report 1,274 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, no clear effect 5%, hurt 17%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add both soft co-annotation and distribution matching | Aff-Wild2 · F1 (expression) | 0.337 -> 0.399 | Helped | Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · 39 citations) |
| Add Maximal Connect quality assessment | Gait3D · Rank-1 | 55.1 -> 55.8 | Helped | QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 2024 · 34 citations) |
| Add KP-RPE | IJB-C unaligned · Verification TAR@0.01% | 21.62 -> 91.85 | Helped | KeyPoint Relative Position Encoding for Face Recognition (CVPR 2024 · 34 citations) |
| add SFAM | PA100K · mA | 81.4 -> 82.9 | Helped | Selective and Orthogonal Feature Activation for Pedestrian Attribute Recognition (AAAI 2024 · 14 citations) |
| Add concept-matrix invariance instead of using feature-invariance alone | DSADS · accuracy (AVG) | 88.6 -> 90.7 | Helped | Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · 12 citations) |
| add RSP perturbations to ArcFace without PTSA | IJB-C (1e-4), MS1MV2 · verification accuracy (%) | 92.52 -> 92.44 | Hurt | TopoFR: A Closer Look at Topology Alignment on Face Recognition (NeurIPS 2024 · 7 citations) |
| Add person location guidance to attribute prediction | VBD group activity retrieval · Hit@1 | 69.5 -> 84.8 | Helped | Learning Group Activity Features Through Person Attribute Prediction (CVPR 2024 · 5 citations) |
| replace the ICT generative encoder with a DeepFill CNN encoder | Synthetic masked LFW UMR-MP · ACC | 97.75% -> 96.42% | Hurt | Masked Face Recognition with Generative-to-Discriminative Representations (ICML 2024 · 5 citations) |
| Use a dynamic l range [-1,1] rather than static l=-1 | CASIA-B mean · Rank-1 accuracy | 94.4 -> 94.7 | Helped | WaveLoss: An Adaptive Dynamic Loss for Deep Gait Recognition (AAAI 2025 · 4 citations) |
| add differential core-template selection to average pooling | IJB-B · TAR@FAR=1e-6 | 38.45 -> 48.01 | Helped | FaceCoresetNet: Differentiable Coresets for Face Set Recognition (AAAI 2024 · 4 citations) |
| Replace the CLIP ResNet-50 visual encoder with a gait-specific DeepGaitV2-3D encoder and add Align and Fusion stages | RA-GAR · F1 | 71.48 -> 76.00 | Helped | RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · 1 citation) |
| Add explicit angular-regression embedding supervision (LAR) to code-prediction loss (LC) | IJB-B · TAR@FAR=1e−4 | Helped | GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 7 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Baselines
- GaitSet · 2 of 7 papers
- GaitGL · 2 of 7 papers
- DeepGaitV2 · 2 of 7 papers
- ArcFace · 2 of 7 papers
- AdaFace · 2 of 7 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 27 papers; from OpenAlex and the papers' affiliations).
- Southern University of Science and Technology · 3 papers
- WATRIX.AI · 3 papers
- Beijing Normal University · 3 papers
- Zhejiang University · 2 papers
- The University of Hong Kong · 2 papers
- Imperial College London · 2 papers
- Chinese Academy of Sciences · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Supervised person re-identification (108 papers)
- Cross-modality and occlusion re-identification (82 papers)
- Fair and quality-aware face recognition (66 papers)
- Uncertainty-aware facial expression recognition (41 papers)
Elsewhere in the topic
- Multilingual language representation pretraining (507 papers)
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).