Research map / Architectures and efficiency / CNN design
Dense matching and estimation CNNs
Designs convolutional descriptors and fusion networks for matching, optical flow, and stereo confidence.
93 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in CNN design, keyword cluster “convolution · kernel · spatial”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
20 papers in 2023–24 and 19 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 14.6% to 18.6%.
Open questions the data points to
- GF2 reduced-resolution (ERGAS): best went from 0.63 (up to 2024) to 0.597 in 2025 — A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · 24 citations)
- WV3 reduced-resolution (SAM): best went from 2.93 (up to 2024) to 2.885 in 2025 — Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations)
What looks settled
- GF2 reduced-resolution (Q4): best up to 2024 0.983, best in 2025–26 0.985 from 2 papers, less than 1% better.
- QB reduced-resolution (SAM): best up to 2024 4.37, best in 2025–26 4.43 from 2 papers, at most 0.2% better, if at all.
- QB reduced-resolution (ERGAS): best up to 2024 3.652, best in 2025–26 3.633 from 2 papers, less than 1% better.
- QB reduced-resolution (Q4): best up to 2024 0.937, best in 2025–26 0.939 from 2 papers, less than 1% better.
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 · Efficient Deep Learning for Stereo Matching (CVPR 2016 · 811 citations)
- 2017 · ECO: Efficient Convolution Operators for Tracking (CVPR 2017 · 2,824 citations)
- 2018 · Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Convolutions (CVPR 2018 · 370 citations)
- 2019 · SiamRPN++: Evolution of Siamese Visual Tracking With Very Deep Networks (CVPR 2019 · 2,558 citations)
- 2020 · Recurrent Feature Reasoning for Image Inpainting (CVPR 2020 · 496 citations)
- 2021 · AdaBins: Depth Estimation Using Adaptive Bins (CVPR 2021 · 717 citations)
- 2022 · Image Dehazing Transformer With Transmission-Aware 3D Position Embedding (CVPR 2022 · 551 citations)
- 2023 · CompletionFormer: Depth Completion With Convolutions and Vision Transformers (CVPR 2023 · 143 citations)
- 2024 · EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · 506 citations)
- 2025 · Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations)
- 2026 · Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models (ICML 2026 · 0 citations)
Earliest papers here
- Efficient Deep Learning for Stereo Matching (CVPR 2016 · 811 citations)
- Learning Sensor Multiplexing Design through Back-propagation (NeurIPS 2016 · 28 citations)
- Efficient Large-Scale Similarity Search Using Matrix Factorization (CVPR 2016 · 19 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.
GF2 reduced-resolution · ERGAS (lower is better)
4 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 0.91 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
| 2024 | 0.63 | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| 2025 | 0.597 | A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · 24 citations) |
GF2 reduced-resolution · Q4 (higher is better)
4 papers, 2022–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 0.97 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
| 2024 | 0.983 | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| 2025 | 0.985 | A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · 24 citations) |
QB reduced-resolution · SAM (lower is better)
4 papers, 2022–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 4.37 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
QB reduced-resolution · ERGAS (lower is better)
4 papers, 2022–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 3.74 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
| 2024 | 3.652 | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| 2025 | 3.633 | Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations) |
QB reduced-resolution · Q4 (higher is better)
4 papers, 2022–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 0.916 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
| 2024 | 0.937 | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| 2025 | 0.939 | Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations) |
WV3 reduced-resolution · SAM (lower is better)
4 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 3.473 | LAGConv: Local-Context Adaptive Convolution Kernels with Global Harmonic Bias for Pansharpening (AAAI 2022 · 118 citations) |
| 2024 | 2.93 | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| 2025 | 2.885 | Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations) |
What has been tried
83 papers here report 958 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 7%, hurt 20%, 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 a cascaded decoder | Synapse Multi-organ · DICE | 80.10±0.2 -> 81.08±0.2 | Helped | EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · 506 citations) |
| Remove synthetic warps from training | Megadepth-1500 · AUC@5° | 42.6 -> 41.5 | Hurt | XFeat: Accelerated Features for Lightweight Image Matching (CVPR 2024 · 185 citations) |
| Add DSC to neither DOR-MLP nor another spatial module | ISIC 2018 (512) · IoU | 79.48 -> 80.62 | Helped | Rolling-Unet: Revitalizing MLP’s Ability to Efficiently Extract Long-Distance Dependencies for Medical Image Segmentation (AAAI 2024 · 138 citations) |
| disable SRP and use K = 1 | WV3 reduced-resolution · SAM | 2.930±0.593 -> 3.023±0.606 | Hurt | Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · 51 citations) |
| Add height and width adaptation | WV3 reduced-resolution · SAM | 2.925±0.593 -> 2.885±0.590 | Helped | Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · 50 citations) |
| use 9×9×9 sparse kernels with SDS and CWS instead of dense 9×9×9 kernels | SemanticKITTI val · mIoU | 67.5 -> 70.2 | Helped | LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · 29 citations) |
| Add three-layer DFU | KITTI Validate Dataset · RMSE [mm] | 772.87 -> 746.98 | Helped | Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · 27 citations) |
| add implicit Retinex-aware guidance instead of estimating the enhanced result directly without a prior | Ablation study · PSNR | 1.25 dB | Helped | MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space (NeurIPS 2024 · 27 citations) |
| Add IFW | WV3 reduced-resolution · PSNR | 38.592 -> 39.028 | Helped | A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · 24 citations) |
| Add masking augmentation to alignment loss (λ=10) | CUB-200-2011 test · PAC | 2.9 ± 1.1 -> 1.7 ± 0.7 | Helped | Interpretability Benchmark for Evaluating Spatial Misalignment of Prototypical Parts Explanations (AAAI 2024 · 10 citations) |
| reduce output resolution from (D, H, W) to (D/2, H/2, W/2) | BTCV validation · average DICE (%) | 79.44% to 79.25% | Hurt | EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation (CVPR 2025 · 9 citations) |
| Increase network depth from N = 1 to N = 2 | CAVE · PSNR | 48.2703 -> 48.7428 | Helped | A Selective Re-learning Mechanism for Hyperspectral Fusion Imaging (CVPR 2025 · 8 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.
Baselines
- MTF-GLP-FS · 2 of 11 papers
- TV · 2 of 11 papers
- PNN · 2 of 11 papers
- PanNet · 2 of 11 papers
- DiCNN · 2 of 11 papers
- FusionNet · 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).
- University of Electronic Science and Technology of China · 3 papers
- Shanghai Jiao Tong University · 2 papers
- University of Texas at Austin · 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
- Efficient CNN scaling and dynamic inference (104 papers)
- Invertible and wide-network analysis (65 papers)
- Large and dynamic kernel design (62 papers)
- Higher-order pooling (20 papers)
Elsewhere in the topic
- Filter design and trained-filter analysis (128 papers)
- Compact CNN architectures (114 papers)
- Theory of convolutional inductive bias (64 papers)
- Group symmetry equivariant networks (55 papers)
- Biologically inspired CNN explanation (46 papers)
- CNN filter pruning (43 papers)
- Steerable rotation-equivariant CNNs (40 papers)
- Group convolution theory (19 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).