Research map / Label-efficient and robust learning / Recognition and long-tail
Correspondence and causal representation learning
Learns visual correspondence and transportable representations by matching multiple representations.
447 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “representations · transfer · unsupervised”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 1 tracked benchmark, but the best result has barely moved since 2024.
98 papers in 2023–24 and 98 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 7.8% to 7.5%.
Open questions the data points to
- Papers report Triplet on CUB200-2011 (Recall@1) anywhere from 35.9 to 49.8: comparisons against it depend on whose number you take.
- Papers report LiftedStruct* on CARS196 (Recall@1) anywhere from 49 to 69.7: comparisons against it depend on whose number you take.
What looks settled
- CIFAR-100 (ACC): best up to 2024 92.6, best in 2025–26 80.86 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: CARS196 (Recall@1, last in 2022); CUB-200-2011 (Recall@1, last in 2023); CUB-200-2011 (NMI, last in 2022); CUB-200-2011 (R@1, last in 2022); Cars196 (R@1, 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 · Deep Metric Learning via Lifted Structured Feature Embedding (CVPR 2016 · 1,740 citations)
- 2017 · L2-Net: Deep Learning of Discriminative Patch Descriptor in Euclidean Space (CVPR 2017 · 622 citations)
- 2018 · Unsupervised Feature Learning via Non-Parametric Instance Discrimination (CVPR 2018 · 3,635 citations)
- 2019 · Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning (CVPR 2019 · 847 citations)
- 2020 · Hyperbolic Image Embeddings (CVPR 2020 · 312 citations)
- 2021 · Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition (CVPR 2021 · 442 citations)
- 2022 · Generalized Category Discovery (CVPR 2022 · 231 citations)
- 2023 · Curricular Contrastive Regularization for Physics-Aware Single Image Dehazing (CVPR 2023 · 336 citations)
- 2024 · Depth Anything V2 (NeurIPS 2024 · 525 citations)
- 2025 · All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 2025 · 7 citations)
- 2026 · RealRep: Generalized SDR-to-HDR Conversion via Attribute-Disentangled Representation Learning (AAAI 2026 · 3 citations)
Earliest papers here
- Deep Metric Learning via Lifted Structured Feature Embedding (CVPR 2016 · 1,740 citations)
- Autoencoding beyond pixels using a learned similarity metric (ICML 2016 · 1,136 citations)
- Joint Unsupervised Learning of Deep Representations and Image Clusters (CVPR 2016 · 802 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.
CARS196 · Recall@1 (higher is better)
14 papers, 2019–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 84.6 | Divide and Conquer the Embedding Space for Metric Learning (CVPR 2019 · 118 citations) |
| 2020 | 85 | Moving in the Right Direction: A Regularization for Deep Metric Learning (CVPR 2020 · 30 citations) |
| 2022 | 91.8 | MetricFormer: A Unified Perspective of Correlation Exploring in Similarity Learning (NeurIPS 2022 · 0 citations) |
CUB-200-2011 · Recall@1 (higher is better)
14 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 45.3 | Mining on Manifolds: Metric Learning Without Labels (CVPR 2018 · 118 citations) |
| 2019 | 67.9 | Towards Optimal Fine Grained Retrieval via Decorrelated Centralized Loss with Normalize-Scale Layer (AAAI 2019 · 78 citations) |
| 2022 | 84 | Hyperbolic Vision Transformers: Combining Improvements in Metric Learning (CVPR 2022 · 113 citations) |
CUB-200-2011 · NMI (higher is better)
9 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 59.23 | Deep Metric Learning via Facility Location (CVPR 2017 · 307 citations) |
| 2019 | 69.6 | Divide and Conquer the Embedding Space for Metric Learning (CVPR 2019 · 118 citations) |
| 2020 | 71.3 | RankMI: A Mutual Information Maximizing Ranking Loss (CVPR 2020 · 45 citations) |
| 2021 | 71.8 | Deep Compositional Metric Learning (CVPR 2021 · 49 citations) |
| 2022 | 75.4 | MetricFormer: A Unified Perspective of Correlation Exploring in Similarity Learning (NeurIPS 2022 · 0 citations) |
CUB-200-2011 · R@1 (higher is better)
9 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 48.18 | Deep Metric Learning via Facility Location (CVPR 2017 · 307 citations) |
| 2019 | 66.1 | Ensemble Deep Manifold Similarity Learning Using Hard Proxies (CVPR 2019 · 63 citations) |
| 2021 | 68.4 | Deep Compositional Metric Learning (CVPR 2021 · 49 citations) |
| 2022 | 71.4 | Hypergraph-Induced Semantic Tuplet Loss for Deep Metric Learning (CVPR 2022 · 44 citations) |
Cars196 · R@1 (higher is better)
9 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 58.11 | Deep Metric Learning via Facility Location (CVPR 2017 · 307 citations) |
| 2019 | 87.6 | Ensemble Deep Manifold Similarity Learning Using Hard Proxies (CVPR 2019 · 63 citations) |
| 2022 | 89.6 | Hypergraph-Induced Semantic Tuplet Loss for Deep Metric Learning (CVPR 2022 · 44 citations) |
CIFAR-100 · ACC (higher is better)
8 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 42.5 | Clustering-friendly Representation Learning via Instance Discrimination and Feature Decorrelation (ICLR 2021 · 36 citations) |
| 2022 | 92.6 | Understanding and Improving Robustness of Vision Transformers through Patch-based Negative Augmentation (NeurIPS 2022 · 13 citations) |
What has been tried
377 papers here report 3,958 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, no clear effect 6%, hurt 15%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add pseudo-labeled real images to synthetic-only student training | DA-2K · accuracy (%) | 89.8 -> 95.3 | Helped | Depth Anything V2 (NeurIPS 2024 · 525 citations) |
| Replace shared coarse/fine encoder weights with separate encoders | Ablation validation · 100-PCK@1px | 17.0 -> 16.0 | Helped | RoMa: Robust Dense Feature Matching (CVPR 2024 · 263 citations) |
| add HFRI | ForenSynths cross-model, 4-class · mean accuracy | 84.3 -> 91.5 | Helped | Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · 193 citations) |
| Fine-tune the final 4 DINOv2 blocks rather than freezing the backbone | MSLS Val · Recall@1 | 88.5 -> 92.2 | Helped | Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · 137 citations) |
| Set skeleton-map Gaussian σ to 1.0 instead of the default 8.0 | Gait3D · rank-1 accuracy | 38.1 -> 34.3 | Hurt | SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · 80 citations) |
| Use a nonlinear MLP head rather than a linear head | survival prediction across TCGA and external cohorts · c-index | Helped | Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology (CVPR 2024 · 70 citations) | |
| Add learnable mask cross-attention | Synapse multi-organ CT, 10% · Mean Dice | 75.57 -> 77.68; 2.1% | Helped | Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · 70 citations) |
| add abstraction identification | ShoeV2 · Acc.@1 | 38.1 -> 45.3 | Helped | How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval? (CVPR 2024 · 34 citations) |
| add CMS learning rather than using embeddings without mean shift | CIFAR100, ImageNet100, CUB, Stanford Cars · All accuracy | 71.5 -> 81.1; 74.1 -> 83.4; 51.2 -> 66.7; 37.9 -> 54.5 | Helped | Contrastive Mean-Shift Learning for Generalized Category Discovery (CVPR 2024 · 33 citations) |
| add LPI, LII, and LPP losses | O&M&I to C · HTER (%) | 17.78 -> 11.30 | Helped | Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · 33 citations) |
| Dynamic Memory Bank added to SSP | nuScenes val ablation · mAP | 0.429 to 0.425 | Hurt | DriveWorld: 4D Pre-trained Scene Understanding via World Models for Autonomous Driving (CVPR 2024 · 29 citations) |
| Use dot product instead of Euclidean distance for similarity measurement | ImageNet-1K val · Top-1 accuracy | 72.2 -> 72.7 | Helped | Neural Clustering based Visual Representation Learning (CVPR 2024 · 26 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Triplet | CUB200-2011 | Recall@1 | 35.9 – 49.8 | 3 papers, e.g. AAAI 2020 and CVPR 2018 |
| LiftedStruct* | CARS196 | Recall@1 | 49 – 69.7 | 3 papers, e.g. AAAI 2019 and CVPR 2019 |
| BigGAN | CIFAR-10 | FID | 6.04 – 16.04 | 2 papers, e.g. ICML 2020 and AAAI 2023 |
| ProPos | ImageNet-Dogs | ACC | 63.9 – 76.9 | 2 papers, e.g. ICML 2025 and NeurIPS 2025 |
| BYOL | CIFAR-10 | Acc% | 83.3 – 98.1 | 2 papers, e.g. ICML 2025 and CVPR 2022 |
| ProPos | Tiny-ImageNet | ACC | 25.6 – 29.8 | 2 papers, e.g. ICML 2025 and NeurIPS 2025 |
What reviewers will expect
Datasets and baselines named in the 51 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-100 · 8 of 51 papers
- CIFAR-10 · 8 of 51 papers
- Tiny-ImageNet · 5 of 51 papers
- ImageNet · 4 of 51 papers
- ImageNet-1K · 4 of 51 papers
- CUB-200-2011 · 3 of 51 papers
- CIFAR-20 · 3 of 51 papers
- STL-10 · 3 of 51 papers
Baselines
- SCAN · 3 of 51 papers
- LoRA · 2 of 51 papers
- CLIP · 2 of 51 papers
- UNet · 2 of 51 papers
- CC · 2 of 51 papers
- GCC · 2 of 51 papers
- TCC · 2 of 51 papers
- IDFD · 2 of 51 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 196 papers; from OpenAlex and the papers' affiliations).
- Chinese Academy of Sciences · 7 papers
- Peking University · 7 papers
- Tencent · 6 papers
- Tsinghua University · 6 papers
- University of Science and Technology of China · 6 papers
- Zhejiang University · 6 papers
- Google · 6 papers
- The Chinese University of Hong Kong · 5 papers
- Shenzhen University · 4 papers
- Huawei · 4 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
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Data augmentation and generalization bias (409 papers)
- Multilingual language representation pretraining (507 papers)
- Recommendation representation learning (207 papers)
Elsewhere in the topic
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
- Learning from unlabeled and positive-unlabeled data (240 papers)
- Confidence calibration objectives (129 papers)
- Margin-based face recognition losses (124 papers)
- Supervised person re-identification (108 papers)
- Self-explaining and visual explanation networks (108 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).