Research map / Detection and segmentation / Open-set and few-shot detection
Open-vocabulary semantic segmentation
Segment arbitrary categories using CLIP, mask adaptation, and diffusion without dense labels.
93 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Open-set and few-shot detection, keyword cluster “vocabulary · language · text”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
30 papers in 2023–24 and 60 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 5.1% to 10.8%.
Open questions the data points to
- ADE20K (mIoU): best went from 30.1 (up to 2024) to 68.4 in 2025 — Scaling up Image Segmentation across Data and Tasks (CVPR 2025 · 1 citation)
- ADE20K A-150 (mIoU): best went from 31.4 (up to 2024) to 41.8 in 2025 — Effective SAM Combination for Open-Vocabulary Semantic Segmentation (CVPR 2025 · 8 citations)
- Papers report SAN on ADE-847 (mIoU) anywhere from 10.1 to 12.4: comparisons against it depend on whose number you take.
- Papers report SAN on ADE-150 (mIoU) anywhere from 27.5 to 32.1: 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
- 2019 · LVIS: A Dataset for Large Vocabulary Instance Segmentation (CVPR 2019 · 1,281 citations)
- 2020 · Two-Level Transformer and Auxiliary Coherence Modeling for Improved Text Segmentation (AAAI 2020 · 42 citations)
- 2022 · Open-Vocabulary Instance Segmentation via Robust Cross-Modal Pseudo-Labeling (CVPR 2022 · 83 citations)
- 2023 · Open-Vocabulary Semantic Segmentation With Mask-Adapted CLIP (CVPR 2023 · 448 citations)
- 2024 · OMG-Seg: Is One Model Good Enough For All Segmentation? (CVPR 2024 · 77 citations)
- 2025 · SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images (CVPR 2025 · 51 citations)
- 2026 · Exploring Efficient Open-Vocabulary Segmentation in the Remote Sensing (AAAI 2026 · 27 citations)
Earliest papers here
- LVIS: A Dataset for Large Vocabulary Instance Segmentation (CVPR 2019 · 1,281 citations)
- Two-Level Transformer and Auxiliary Coherence Modeling for Improved Text Segmentation (AAAI 2020 · 42 citations)
- Open-Vocabulary Instance Segmentation via Robust Cross-Modal Pseudo-Labeling (CVPR 2022 · 83 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.
ADE20K · mIoU (higher is better)
6 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 17.1 | Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs (CVPR 2023 · 90 citations) |
| 2024 | 30.1 | CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction (ICLR 2024 · 12 citations) |
| 2025 | 68.4 | Scaling up Image Segmentation across Data and Tasks (CVPR 2025 · 1 citation) |
Pascal VOC · mIoU (higher is better)
6 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 94.6 | Side Adapter Network for Open-Vocabulary Semantic Segmentation (CVPR 2023 · 308 citations) |
| 2024 | 97.2 | Open-Vocabulary Segmentation with Semantic-Assisted Calibration (CVPR 2024 · 40 citations) |
ADE20K A-150 · mIoU (higher is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 31.4 | Open Vocabulary Semantic Segmentation With Patch Aligned Contrastive Learning (CVPR 2023 · 72 citations) |
| 2025 | 41.8 | Effective SAM Combination for Open-Vocabulary Semantic Segmentation (CVPR 2025 · 8 citations) |
COCO Object · mIoU (higher is better)
5 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 31.6 | Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs (CVPR 2023 · 90 citations) |
| 2024 | 44.5 | Image-to-Image Matching via Foundation Models: A New Perspective for Open-Vocabulary Semantic Segmentation (CVPR 2024 · 16 citations) |
PASCAL Context · mIoU (higher is better)
5 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 30.4 | Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs (CVPR 2023 · 90 citations) |
| 2024 | 62.3 | CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction (ICLR 2024 · 12 citations) |
PASCAL Context val · mIoU (higher is better)
5 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 24.7 | SegCLIP: Patch Aggregation with Learnable Centers for Open-Vocabulary Semantic Segmentation (ICML 2023 · 27 citations) |
| 2024 | 37.9 | Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels (NeurIPS 2024 · 8 citations) |
What has been tried
55 papers here report 689 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 3%, hurt 21%, 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 VIPSeg and YT-VIS-19 training data to COCO co-training | VIPSeg-VPS · VPQ | 32.2 -> 49.0 | Helped | OMG-Seg: Is One Model Good Enough For All Segmentation? (CVPR 2024 · 77 citations) |
| Upsample earlier-stage CLIP features | OpenEarthMap · mIoU | 32.4 -> 34.6 | Helped | SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images (CVPR 2025 · 51 citations) |
| add Semantic Integration Module | ADE20K-150 · mIoU | 31.5 -> 32.9 | Helped | Open-Vocabulary Segmentation with Semantic-Assisted Calibration (CVPR 2024 · 40 citations) |
| use recurrent refinement instead of a single pass | Pascal VOC · mIoU | 15.2 -> 67.6 | Helped | CLIP as RNN: Segment Countless Visual Concepts without Training Endeavor (CVPR 2024 · 34 citations) |
| Add the AdaB Decoder and SSC Loss | ADE20K-150 · mIoU | 29.2 -> 30.0 | Helped | Open-Vocabulary Semantic Segmentation with Image Embedding Balancing (CVPR 2024 · 28 citations) |
| Add global CLIP similarity to local DINOv2 ViT-B/14 similarity | Pascal VOC · mIoU | 75.6 -> 86.2 | Helped | Training-Free Open-Vocabulary Segmentation with Offline Diffusion-Augmented Prototype Generation (CVPR 2024 · 28 citations) |
| Use LVLM to identify concrete entities from images instead of extracting nouns from captions | ScanNet val, annotation-free entity identification · mIoU | 55.3 -> 59.6 | Helped | Open-Vocabulary 3D Semantic Segmentation with Foundation Models (CVPR 2024 · 20 citations) |
| Add ground-truth mask warmup | ADE20K · mIoU | 34.1 -> 34.3 | Helped | Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation (CVPR 2025 · 18 citations) |
| Combine CLIP and DINOv2 instead of using CLIP only | ADE-150 · mIoU | 30.2 -> 32.6 | Helped | USE: Universal Segment Embeddings for Open-Vocabulary Image Segmentation (CVPR 2024 · 17 citations) |
| replace CLIP image-to-text matching with DINOv2 image-to-image matching | COCO Object · mIoU | 26.7 -> 38.7 | Helped | Image-to-Image Matching via Foundation Models: A New Perspective for Open-Vocabulary Semantic Segmentation (CVPR 2024 · 16 citations) |
| Use patch-split maximum M=6 instead of M=4, M=8, or M=10 | COCO val2017 region boxes · Top1 mAcc | 71.3, 71.6, and 70.0 -> 72.1 | Helped | CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction (ICLR 2024 · 12 citations) |
| Add spectral VFM distillation with graph matching, low-rank approximation, and dynamic eigenscaling | PASCAL VOC 2012 with background (V21) · mIoU | 58.9 -> 64.8 | Helped | Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation (CVPR 2025 · 11 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 |
|---|---|---|---|---|
| SAN | ADE-847 | mIoU | 10.1 – 12.4 | 3 papers, e.g. AAAI 2025 and CVPR 2024 |
| SAN | ADE-150 | mIoU | 27.5 – 32.1 | 3 papers, e.g. AAAI 2025 and CVPR 2024 |
| OpenSeg | ADE20K (A-150) | mIoU | 21.1 – 28.6 | 2 papers, e.g. ICML 2023 and CVPR 2023 |
| CLIPpy | COCO Object | mIoU | 25.5 – 32 | 2 papers, e.g. CVPR 2023 and CVPR 2024 |
| GroupViT | PASCAL Context val | mIoU | 18.2 – 22.4 | 2 papers, e.g. NeurIPS 2023 and ICML 2023 |
| ODISE | COCO Panoptic | PQ | 45.3 – 55.4 | 2 papers, e.g. ICLR 2024 and NeurIPS 2023 |
What reviewers will expect
Datasets and baselines named in the 31 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-Stuff · 7 of 31 papers
- COCO · 5 of 31 papers
- PASCAL Context · 5 of 31 papers
- ADE20K · 5 of 31 papers
- LVIS · 4 of 31 papers
- PASCAL VOC 2012 · 3 of 31 papers
- Cityscapes · 3 of 31 papers
- RefCOCO · 3 of 31 papers
Baselines
- MaskCLIP · 7 of 31 papers
- CAT-Seg · 7 of 31 papers
- OVSeg · 6 of 31 papers
- SAN · 5 of 31 papers
- TCL · 5 of 31 papers
- SED · 5 of 31 papers
- ZegFormer · 5 of 31 papers
- EBSeg · 4 of 31 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 90 papers; from OpenAlex and the papers' affiliations).
- Shanghai Jiao Tong University · 4 papers
- Shanghai AI Laboratory · 4 papers
- Nanyang Technological University · 4 papers
- The Chinese University of Hong Kong · 4 papers
- ByteDance · 3 papers
- University of Oxford · 3 papers
- Yonsei University · 3 papers
- University of Chinese Academy of Sciences · 2 papers
- Amazon · 2 papers
- Fudan University · 2 papers
Use it
Download BibTeX of 33 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-vocabulary object detection (96 papers)
- Open-vocabulary 3D and world perception (41 papers)
Elsewhere in the topic
- Object-level representation learning (219 papers)
- Interactive and referring segmentation (161 papers)
- Open-set and unknown object detection (160 papers)
- Amodal and open-world instance segmentation (137 papers)
- Industrial anomaly detection (117 papers)
- Unsupervised object and part discovery (114 papers)
- Domain-adaptive object detection (103 papers)
- Object-centric learning theory and video (86 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).