Research map / Detection and segmentation / CNN features
Convolution-transformer backbones
Design visual backbones mixing convolution and self-attention, including masked pretraining.
101 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in CNN features, keyword cluster “vision · imagenet · transformers”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 5 tracked benchmarks.
21 papers in 2023–24 and 21 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 6.4% to 6.1%.
Open questions the data points to
- ADE20K (mIoU): best went from 50 (up to 2024) to 52.1 in 2025 — VFM-Adapter: Adapting Visual Foundation Models for Dense Prediction with Dynamic Hybrid Operation Mapping (AAAI 2025 · 1 citation)
No longer reported by papers here: PASCAL VOC 2012 test (mIoU, 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 Residual Learning for Image Recognition (CVPR 2016 · 228,971 citations)
- 2017 · Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs (CVPR 2017 · 1,859 citations)
- 2018 · MobileNetV2: Inverted Residuals and Linear Bottlenecks (CVPR 2018 · 26,804 citations)
- 2019 · Deformable ConvNets V2: More Deformable, Better Results (CVPR 2019 · 2,774 citations)
- 2020 · RandAugment: Practical Automated Data Augmentation with a Reduced Search Space (NeurIPS 2020 · 3,053 citations)
- 2021 · Rethinking Semantic Segmentation From a Sequence-to-Sequence Perspective With Transformers (CVPR 2021 · 3,682 citations)
- 2022 · SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation (NeurIPS 2022 · 488 citations)
- 2023 · PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers (AAAI 2023 · 115 citations)
- 2024 · PTQ4SAM: Post-Training Quantization for Segment Anything (CVPR 2024 · 27 citations)
- 2025 · OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels (CVPR 2025 · 93 citations)
- 2026 · LWGANet: Addressing Spatial and Channel Redundancy in Remote Sensing Visual Tasks with Light-Weight Grouped Attention (AAAI 2026 · 21 citations)
Earliest papers here
- Deep Residual Learning for Image Recognition (CVPR 2016 · 228,971 citations)
- Online Detection and Classification of Dynamic Hand Gestures With Recurrent 3D Convolutional Neural Network (CVPR 2016 · 706 citations)
- Object Detection From Video Tubelets With Convolutional Neural Networks (CVPR 2016 · 282 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)
7 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 79.2 | Improving Convolutional Networks With Self-Calibrated Convolutions (CVPR 2020 · 503 citations) |
| 2023 | 88.3 | PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers (AAAI 2023 · 115 citations) |
ADE20K · mIoU (higher is better)
6 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 39.33 | Bag of Tricks for Image Classification with Convolutional Neural Networks (CVPR 2019 · 1,582 citations) |
| 2022 | 50 | Image BERT Pre-training with Online Tokenizer (ICLR 2022 · 205 citations) |
| 2025 | 52.1 | VFM-Adapter: Adapting Visual Foundation Models for Dense Prediction with Dynamic Hybrid Operation Mapping (AAAI 2025 · 1 citation) |
ImageNet-1K val · top-1 accuracy (higher is better)
6 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 79.3 | Tied Block Convolution: Leaner and Better CNNs with Shared Thinner Filters (AAAI 2021 · 46 citations) |
| 2022 | 86.5 | Focal Modulation Networks (NeurIPS 2022 · 146 citations) |
ADE20K val · mIoU (higher is better)
4 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 25.9 | Building Detail-Sensitive Semantic Segmentation Networks With Polynomial Pooling (CVPR 2019 · 28 citations) |
| 2021 | 50.28 | Rethinking Semantic Segmentation From a Sequence-to-Sequence Perspective With Transformers (CVPR 2021 · 3,682 citations) |
| 2023 | 50.67 | 1% VS 100%: Parameter-Efficient Low Rank Adapter for Dense Predictions (CVPR 2023 · 45 citations) |
PASCAL VOC 2012 test · mIoU (higher is better)
4 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 74.8 | Dual Path Networks (NeurIPS 2017 · 446 citations) |
| 2019 | 85.6 | Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image Segmentation (CVPR 2019 · 1,058 citations) |
| 2022 | 90.6 | SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation (NeurIPS 2022 · 488 citations) |
What has been tried
85 papers here report 1,034 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 7%, hurt 18%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace vanilla convolution with Dilated RepConv | ImageNet-1K (120 epochs) · Top-1 accuracy | 76.3% to 76.6% | Helped | OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels (CVPR 2025 · 93 citations) |
| add Bimodal Integration and Adaptive Granularity Quantization | COCO instance segmentation · mAP | 25.8 -> 32.1 | Helped | PTQ4SAM: Post-Training Quantization for Segment Anything (CVPR 2024 · 27 citations) |
| truncate LLaMA 7B to 6 decoder blocks instead of retaining 32 | CC3M · F1 | 0.632 -> 0.627 | No clear effect | Object Recognition as Next Token Prediction (CVPR 2024 · 7 citations) |
| add the local branch to the global branch | ImageNet-1K classification · top-1 accuracy | 78.0 -> 82.5 | Helped | Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · 5 citations) |
| Replace the non-overlapping patchify stem with overlapping convolutions | ImageNet-1K · Top-1 accuracy | 74.9% -> 76.7% | Helped | IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION (ICLR 2025 · 4 citations) |
| Use semantic-aware order instead of raster order | ImageNet classification · top-1 accuracy | 82.9 -> 83.5 | Helped | Semantic-Aware Autoregressive Image Modeling for Visual Representation Learning (AAAI 2024 · 4 citations) |
| Denoise only final-layer features instead of each layer | Market1501 · mAP | 91.07% vs 92.05% | Hurt | DenoiseRep: Denoising Model for Representation Learning (NeurIPS 2024 · 3 citations) |
| use probability sorting instead of bidirectional scanning | GlySAC · miscellaneous nuclei F1 | 29.27 -> 35.92 | Helped | Category Prompt Mamba Network for Nuclei Segmentation and Classification (AAAI 2025 · 2 citations) |
| Add G-Biases to SRE (C2) | CIFAR-10 · Top-1 accuracy | 94.4 -> 95.0 (+0.6) | Helped | Relaxed Rotational Equivariance via G-Biases in Vision (AAAI 2025 · 2 citations) |
| Use parameter-inverted rather than parameter-direct branch sizes | COCO val2017, Mask R-CNN 1× · AP^b | 42.6 -> 46.6 | Helped | Parameter-Inverted Image Pyramid Networks (NeurIPS 2024 · 2 citations) |
| Add dynamic clustering convolution | ImageNet-1K · top-1 accuracy | 75.3% -> 77.4% | Helped | Dynamic Clustering Convolutional Neural Network (AAAI 2025 · 1 citation) |
| Add hybrid operation mapping | COCO ATSS · bbox mAP | 40.1 -> 42.2 | Helped | VFM-Adapter: Adapting Visual Foundation Models for Dense Prediction with Dynamic Hybrid Operation Mapping (AAAI 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 13 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 · 7 of 13 papers
- ImageNet-1K · 6 of 13 papers
- COCO 2017 · 4 of 13 papers
- COCO · 2 of 13 papers
Baselines
- ConvNeXt · 3 of 13 papers
- Swin Transformer · 2 of 13 papers
- VMamba · 2 of 13 papers
- UperNet · 2 of 13 papers
- Mask2Former · 2 of 13 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 42 papers; from OpenAlex and the papers' affiliations).
- University of Chinese Academy of Sciences · 5 papers
- Chinese Academy of Sciences · 4 papers
- University of Science and Technology of China · 3 papers
- Megvii · 3 papers
- Tsinghua University · 3 papers
- Peking University · 3 papers
- Shanghai Jiao Tong University · 3 papers
- SenseTime · 3 papers
- Meta · 2 papers
- ByteDance · 2 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
- Plain and high-resolution ViTs (69 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).