Research map / Detection and segmentation / Open-set and few-shot detection
Object-level representation learning
Learn object-centric recognition features through contrastive pretraining and context debiasing.
219 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Open-set and few-shot detection, keyword cluster “images · anomaly · interactive”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
61 papers in 2023–24 and 38 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 10.4% to 6.9%.
Open questions the data points to
- Papers report MAE on ADE20K (mIoU) anywhere from 45.9 to 51.5: comparisons against it depend on whose number you take.
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 · Generation and Comprehension of Unambiguous Object Descriptions (CVPR 2016 · 1,214 citations)
- 2017 · Physically-Based Rendering for Indoor Scene Understanding Using Convolutional Neural Networks (CVPR 2017 · 281 citations)
- 2018 · COCO-Stuff: Thing and Stuff Classes in Context (CVPR 2018 · 1,318 citations)
- 2019 · Character Region Awareness for Text Detection (CVPR 2019 · 1,103 citations)
- 2020 · Interpretable and Accurate Fine-grained Recognition via Region Grouping (CVPR 2020 · 159 citations)
- 2021 · Distribution Alignment: A Unified Framework for Long-Tail Visual Recognition (CVPR 2021 · 300 citations)
- 2022 · Fine-Grained Object Classification via Self-Supervised Pose Alignment (CVPR 2022 · 109 citations)
- 2023 · Universal Instance Perception As Object Discovery and Retrieval (CVPR 2023 · 140 citations)
- 2024 · Composing Object Relations and Attributes for Image-Text Matching (CVPR 2024 · 43 citations)
- 2025 · 5%>100%: Breaking Performance Shackles of Full Fine-Tuning on Visual Recognition Tasks (CVPR 2025 · 70 citations)
- 2026 · CountGD++: Generalized Prompting for Open-World Counting (CVPR 2026 · 1 citation)
Earliest papers here
- Generation and Comprehension of Unambiguous Object Descriptions (CVPR 2016 · 1,214 citations)
- Deep Interactive Object Selection (CVPR 2016 · 423 citations)
- Picking Deep Filter Responses for Fine-Grained Image Recognition (CVPR 2016 · 338 citations)
What has been tried
196 papers here report 2,071 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 8%, hurt 17%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use intermediate dimension 64 instead of 32 or 128 | Pascal VOC · APBox | 86.8 %; 87.3 %; 87.1 % | Mixed | 5%>100%: Breaking Performance Shackles of Full Fine-Tuning on Visual Recognition Tasks (CVPR 2025 · 70 citations) |
| Use the two-step Obj-Att & Obj-Obj GAT encoding instead of one joint GAT | Flickr30K test · RSUM | 504.8 -> 515.8 | Helped | Composing Object Relations and Attributes for Image-Text Matching (CVPR 2024 · 43 citations) |
| pretrain CLIP-UNet on LAION-SP before HSSD training | HSSD VAL · TrP | 16.2 -> 18.5 | Helped | Seeing the Unseen: Visual Common Sense for Semantic Placement (CVPR 2024 · 16 citations) |
| Use SAM ViT-B backbone rather than GES-GSP baseline | StitchBench · MDR | 1.9% | Helped | Object-level Geometric Structure Preserving for Natural Image Stitching (AAAI 2025 · 14 citations) |
| Add quality-aware multi-assignment | Visual Genome · R@100 | 32.1 -> 33.1 | Helped | Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (CVPR 2024 · 13 citations) |
| Add the RoI loss | Concept transfer · concept geometric and structural detail | Helped | CLiC: Concept Learning in Context (CVPR 2024 · 12 citations) | |
| Add a verification stage by using EVE instead of EE | HM3D-SEM validation · Success | 0.590 -> 0.666 | Helped | Instance-aware Exploration-Verification-Exploitation for Instance ImageGoal Navigation (CVPR 2024 · 12 citations) |
| remove unique-assignment and equipartition constraints | CUB-200-2011 · classification accuracy (%) | 91.02 -> 9.25 | Hurt | Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · 11 citations) |
| Replace SAC with DDPG | Light-controlled darkroom · reward per frame | 1.61 -> 1.11 | Hurt | Learning to Control Camera Exposure via Reinforcement Learning (CVPR 2024 · 11 citations) |
| Add the Collaboration Score to the Clue Score | MS COCO · Insertion | 0.3632 -> 0.5459 | Helped | Interpreting Object-level Foundation Models via Visual Precision Search (CVPR 2025 · 10 citations) |
| use the class token rather than a random embedding as automatic saliency guidance | AIM-500 · SAD | 413.86 -> 34.30 | Helped | Unifying Automatic and Interactive Matting with Pretrained ViTs (CVPR 2024 · 9 citations) |
| combine image tags with the text query in the inpainting prompt | Overall evaluation dataset · LPIPS | 0.327 -> 0.320 | Helped | Open-World Amodal Appearance Completion (CVPR 2025 · 8 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.
What reviewers will expect
Datasets and baselines named in the 26 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- COCO · 4 of 26 papers
- PASCAL VOC 2012 · 2 of 26 papers
- Objects365 · 2 of 26 papers
- ADE20K · 2 of 26 papers
- RefCOCO · 2 of 26 papers
- RefCOCOg · 2 of 26 papers
- Visual Genome · 2 of 26 papers
- cue conflict · 2 of 26 papers
Baselines
- Grounding-DINO · 3 of 26 papers
- DetDiffusion · 2 of 26 papers
- GLIP T · 2 of 26 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 99 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 6 papers
- Meta · 6 papers
- Harbin Institute of Technology · 4 papers
- Shanghai Jiao Tong University · 4 papers
- The University of Hong Kong · 4 papers
- Peking University · 3 papers
- Amazon · 3 papers
- Georgia Institute of Technology · 3 papers
- National University of Singapore · 3 papers
- Tencent · 3 papers
Use it
Download BibTeX of 40 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
- Open-set and unknown object detection (160 papers)
- Interactive and referring segmentation (161 papers)
- Amodal and open-world instance segmentation (137 papers)
- Industrial anomaly detection (117 papers)
- Remote sensing segmentation and change (84 papers)
Elsewhere in the topic
- Unsupervised object and part discovery (114 papers)
- Domain-adaptive object detection (103 papers)
- Open-vocabulary object detection (96 papers)
- Open-vocabulary semantic segmentation (93 papers)
- Object-centric learning theory and video (86 papers)
- Few-shot object detection (67 papers)
- Human-object interaction detection (66 papers)
- Few-shot segmentation (64 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).