Research map / Detection and segmentation / CNN features
Local features and context in CNNs
Learn local descriptors and contextual relations to aid detection and recognition.
170 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in CNN features, keyword cluster “local · context · semantic segmentation”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
20 papers in 2023–24 and 16 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.1% to 4.6%.
No longer reported by papers here: CUB-200-2011 (accuracy, last in 2022); Stanford Cars (accuracy, last in 2020); Cityscapes val (mIoU, last in 2023); FGVC-Aircraft (accuracy, last in 2020); PASCAL VOC 2007 test (mAP, last in 2017).
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 · Learning Deep Features for Discriminative Localization (CVPR 2016 · 11,042 citations)
- 2017 · Pyramid Scene Parsing Network (CVPR 2017 · 15,903 citations)
- 2018 · Relation Networks for Object Detection (CVPR 2018 · 1,559 citations)
- 2019 · D2-Net: A Trainable CNN for Joint Description and Detection of Local Features (CVPR 2019 · 1,184 citations)
- 2020 · Random Erasing Data Augmentation (AAAI 2020 · 2,930 citations)
- 2021 · Dual-level Collaborative Transformer for Image Captioning (AAAI 2021 · 323 citations)
- 2022 · Dual Cross-Attention Learning for Fine-Grained Visual Categorization and Object Re-Identification (CVPR 2022 · 277 citations)
- 2023 · The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector (CVPR 2023 · 79 citations)
- 2024 · Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (CVPR 2024 · 36 citations)
- 2025 · Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · 14 citations)
- 2026 · WorldComp2D: Spatio-semantic Representations of Object Identity and Location from Local Views (ICML 2026 · 0 citations)
Earliest papers here
- Learning Deep Features for Discriminative Localization (CVPR 2016 · 11,042 citations)
- Context Encoders: Feature Learning by Inpainting (CVPR 2016 · 5,770 citations)
- Inside-Outside Net: Detecting Objects in Context With Skip Pooling and Recurrent Neural Networks (CVPR 2016 · 1,326 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.
CUB-200-2011 · accuracy (higher is better)
9 papers, 2016–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 76.2 | Part-Stacked CNN for Fine-Grained Visual Categorization (CVPR 2016 · 479 citations) |
| 2017 | 85.8 | Deep Co-Occurrence Feature Learning for Visual Object Recognition (CVPR 2017 · 46 citations) |
| 2019 | 87.9 | Looking for the Devil in the Details: Learning Trilinear Attention Sampling Network for Fine-Grained Image Recognition (CVPR 2019 · 499 citations) |
| 2020 | 89.09 | Filtration and Distillation: Enhancing Region Attention for Fine-Grained Visual Categorization (AAAI 2020 · 199 citations) |
Stanford Cars · accuracy (higher is better)
5 papers, 2017–2020 · not reported since 2020
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 92.5 | Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-Grained Image Recognition (CVPR 2017 · 1,477 citations) |
| 2019 | 94.5 | Destruction and Construction Learning for Fine-Grained Image Recognition (CVPR 2019 · 556 citations) |
Cityscapes val · mIoU (higher is better)
4 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 76.4 | Spatial as Deep: Spatial CNN for Traffic Scene Understanding (AAAI 2018 · 1,120 citations) |
| 2021 | 80.7 | Generic Perceptual Loss for Modeling Structured Output Dependencies (CVPR 2021 · 48 citations) |
| 2023 | 87.1 | DejaVu: Conditional Regenerative Learning To Enhance Dense Prediction (CVPR 2023 · 8 citations) |
FGVC-Aircraft · accuracy (higher is better)
4 papers, 2016–2020 · not reported since 2020
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 88.4 | Mining Discriminative Triplets of Patches for Fine-Grained Classification (CVPR 2016 · 124 citations) |
| 2019 | 93 | Destruction and Construction Learning for Fine-Grained Image Recognition (CVPR 2019 · 556 citations) |
| 2020 | 93.37 | Filtration and Distillation: Enhancing Region Attention for Fine-Grained Visual Categorization (AAAI 2020 · 199 citations) |
PASCAL VOC 2007 test · mAP (higher is better)
4 papers, 2016–2017 · not reported since 2017
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 92.9 | Region Ranking SVM for Image Classification (CVPR 2016 · 23 citations) |
What has been tried
158 papers here report 1,734 measured changes (a component added, removed or swapped, compared with a baseline): helped 77%, no clear effect 5%, hurt 16%, 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 IAKD with FPS keypoint sampling | REAL275 · 5°2 cm | 54.7 -> 46.2 | Hurt | Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (CVPR 2024 · 36 citations) |
| Add simple re-ranking | MSLS-challenge · R@1 | 74.8 -> 75.7 | Helped | Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · 14 citations) |
| increase Tmax from 5 to 10 | MixATIS · Overall Acc | 51.9 -> 52.5 | Helped | Towards Explainable Joint Models via Information Theory for Multiple Intent Detection and Slot Filling (AAAI 2024 · 12 citations) |
| Remove the guidance-signal MLP | ADE20K val · mIoU | 44.2 -> 42.9 | Hurt | FeatUp: A Model-Agnostic Framework for Features at Any Resolution (ICLR 2024 · 9 citations) |
| Add LARL | SPair-71K, αbbox=0.1 · PCK | 63.2 -> 65.6 | Helped | Pixel-level Semantic Correspondence through Layout-aware Representation Learning and Multi-scale Matching Integration (CVPR 2024 · 7 citations) |
| replace add fusion with guided cross-attention using the image feature as query | MatterportLayout · 3D IoU | 81.26 -> 82.57 | Helped | No More Ambiguity in 360deg Room Layout via Bi-Layout Estimation (CVPR 2024 · 6 citations) |
| use recurrent support feature mining | MP-100 Split1 1-shot · mPCK | 69.82 -> 73.18 (+3.35% mPCK) | Helped | Recurrent Feature Mining and Keypoint Mixup Padding for Category-Agnostic Pose Estimation (CVPR 2025 · 3 citations) |
| Add semantic fusion (instead of split&concat) | NFL test · R10@1 | 42.74 -> 45.61 | Helped | Tensorized Attention for Understanding Multi-Object Relationships (AAAI 2025 · 1 citation) |
| Train for 25 rather than 20 epochs | VideoMatte240K-JPEG-SD test (224×224) · MAD | 4.1357 -> 4.1811 | Hurt | EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting (CVPR 2024 · 1 citation) |
| Remove weak-to-strong learning | ImageNet-C severity 5 · accuracy | 70.1 -> 4.9 | Hurt | Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · 0 citations) |
| Add spatial context understanding (SCU) | Simulation Moderate · success rate (%) | 76.1 -> 81.0 | Helped | Learning Spatial-Aware Manipulation Ordering (NeurIPS 2025 · 0 citations) |
| Use SLIC prompting instead of grid prompting | ADE20K · mAP | 50.4 -> 50.9 | Helped | REN: Fast and Efficient Region Encodings from Patch-Based Image Encoders (NeurIPS 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.
Baselines
- GPT-4o · 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 36 papers; from OpenAlex and the papers' affiliations).
- Tencent · 4 papers
- Shanghai Jiao Tong University · 3 papers
- Peking University · 3 papers
- University of Science and Technology of China · 3 papers
- Chinese Academy of Sciences · 2 papers
- Fudan University · 2 papers
- South China University of Technology · 2 papers
- SenseTime · 2 papers
- Tsinghua University · 2 papers
- The University of Sydney · 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
- Context modeling for semantic segmentation (241 papers)
- Context for hard detection conditions (133 papers)
- Domain-specific segmentation and detection (118 papers)
- Forgery detection (46 papers)
- Lane detection (32 papers)
Elsewhere in the topic
- Region-based and query-based detectors (183 papers)
- Convolution-transformer backbones (101 papers)
- Box regression and loss design (94 papers)
- Efficient DETR query design (85 papers)
- Boundary-aware instance segmentation (83 papers)
- Plain and high-resolution ViTs (69 papers)
- Arbitrary-shape scene text detection (61 papers)
- RGB salient object detection (61 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).