Research map / Detection and segmentation / CNN features
Plain and high-resolution ViTs
Adapt vision transformers to segmentation and detection with high-resolution and plain designs.
69 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in CNN features, keyword cluster “vision · imagenet · transformers”
State History Benchmarks Tried 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.
22 papers in 2023–24 and 17 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.7% to 4.9%.
What looks settled
- ADE20K (mIoU): best up to 2024 51.8, best in 2025–26 49.1 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: Cityscapes val (mIoU, last in 2023); ADE20K val (mIoU, last in 2023).
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
- 2021 · SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (NeurIPS 2021 · 914 citations)
- 2022 · A ConvNet for the 2020s (CVPR 2022 · 8,138 citations)
- 2023 · Vision Transformer Adapter for Dense Predictions (ICLR 2023 · 199 citations)
- 2024 · TransNeXt: Robust Foveal Visual Perception for Vision Transformers (CVPR 2024 · 401 citations)
- 2025 · MambaVision: A Hybrid Mamba-Transformer Vision Backbone (CVPR 2025 · 308 citations)
- 2026 · Adaptive Morph-Patch Transformer for Aortic Vessel Segmentation (AAAI 2026 · 2 citations)
Earliest papers here
- SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (NeurIPS 2021 · 914 citations)
- XCiT: Cross-Covariance Image Transformers (NeurIPS 2021 · 231 citations)
- You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection (NeurIPS 2021 · 192 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.
ImageNet-1K · Top-1 accuracy (higher is better)
8 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 84.3 | MPViT: Multi-Path Vision Transformer for Dense Prediction (CVPR 2022 · 357 citations) |
| 2023 | 85.7 | Global Context Vision Transformers (ICML 2023 · 36 citations) |
ImageNet-1K val · top-1 accuracy (%) (higher is better)
8 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 84 | Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · 75 citations) |
| 2022 | 85.4 | Orthogonal Transformer: An Efficient Vision Transformer Backbone with Token Orthogonalization (NeurIPS 2022 · 2 citations) |
ADE20K · mIoU (higher is better)
6 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 51.8 | SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (NeurIPS 2021 · 914 citations) |
Cityscapes val · mIoU (higher is better)
5 papers, 2021–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 84 | SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (NeurIPS 2021 · 914 citations) |
ADE20K val · mIoU (higher is better)
4 papers, 2022–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 59.9 | Swin Transformer V2: Scaling Up Capacity and Resolution (CVPR 2022 · 2,405 citations) |
What has been tried
59 papers here report 740 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 7%, hurt 21%, 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 |
|---|---|---|---|---|
| Replace ConvFFN with GLU | ImageNet-1K val, ablation roadmap · top-1 accuracy | 80.12 -> 79.7 (-0.42) | Hurt | TransNeXt: Robust Foveal Visual Perception for Vision Transformers (CVPR 2024 · 401 citations) |
| Replace causal conv1 with regular convolution | ImageNet-1K · Top-1 accuracy | 80.5 -> 80.9 | Helped | MambaVision: A Hybrid Mamba-Transformer Vision Backbone (CVPR 2025 · 308 citations) |
| Use MAEBBoxHead without pre-training | DOTA-v1.0 · mAP | 71.07 -> 69.16 | Hurt | Spatial Transform Decoupling for Oriented Object Detection (AAAI 2024 · 54 citations) |
| Replace HAT with Twins attention | ImageNet-1K val · Top-1 accuracy (%) | 80.8 -> 82.1 | Hurt | FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · 35 citations) |
| share all decoder components across one-to-one and one-to-many tasks | COCO 2017 val, 12 ep, 900 queries · mAP (o2o) | 47.6 -> 41.6 | Hurt | Mr. DETR: Instructive Multi-Route Training for Detection Transformers (CVPR 2025 · 16 citations) |
| increase computational-cost penalty λc to 1e-2 | LOD · total time (ms) | 14.73 -> 11.54 | Helped | AdaptiveISP: Learning an Adaptive Image Signal Processor for Object Detection (NeurIPS 2024 · 10 citations) |
| use 16-head MHA with QKV and head projections and LayerNorm instead of simple single-head cross-attention | ImageNet-1k · accuracy | 85.3 -> 85.2 | Hurt | Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · 3 citations) |
| Use dynamic upsampling to match the encoder embedding size instead of a fixed-size resize | COCO · AP | 0.5 AP drop | Helped | DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · 1 citation) |
| Use 3 heads rather than 5 | ADE20K · mIoU | 45.6 -> 45.6 | No clear effect | Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · 1 citation) |
| Add learnable mask blocks in the first three encoder stages | BDD validation (trained on Cityscapes) · mIoU | 46.7 -> 48.1 | Helped | MaskViM: Domain Generalized Semantic Segmentation with State Space Models (AAAI 2025 · 0 citations) |
| add the 2D polyline path mask | ImageNet-1K · Top-1 accuracy | +0.32% | Helped | Polyline Path Masked Attention for Vision Transformer (NeurIPS 2025 · 0 citations) |
| increase PRO scaling configuration from (1,1,1) to (3,3,3) | COCO val · PQ | 50.79 -> 52.82 | Helped | Progressive Token Length Scaling in Transformer Encoders for Efficient Universal Segmentation (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 11 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ADE20K · 6 of 11 papers
- ImageNet-1K · 5 of 11 papers
- Cityscapes · 3 of 11 papers
- COCO 2017 · 3 of 11 papers
- COCO · 3 of 11 papers
Baselines
- VMamba · 3 of 11 papers
- Mask2former · 2 of 11 papers
- ConvNeXt · 2 of 11 papers
- Deformable-DETR++ · 2 of 11 papers
- DINO · 2 of 11 papers
- Lite DETR · 2 of 11 papers
- Focus DETR · 2 of 11 papers
- DETR · 2 of 11 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 39 papers; from OpenAlex and the papers' affiliations).
- NVIDIA · 4 papers
- Fudan University · 3 papers
- Sogang University · 2 papers
- Eindhoven University of Technology · 2 papers
- Shanghai AI Laboratory · 2 papers
- Tencent · 2 papers
- Peking University · 2 papers
Use it
Download BibTeX of 38 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
- Convolution-transformer backbones (101 papers)
- Efficient network training and design (37 papers)
- Vision state space models (13 papers)
Elsewhere in the topic
- Context modeling for semantic segmentation (241 papers)
- Region-based and query-based detectors (183 papers)
- Local features and context in CNNs (170 papers)
- Context for hard detection conditions (133 papers)
- Domain-specific segmentation and detection (118 papers)
- Box regression and loss design (94 papers)
- Efficient DETR query design (85 papers)
- Boundary-aware instance segmentation (83 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).