Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly supervised detection
Train detectors from points, videos, or image labels, with weak-to-full frameworks.
97 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Weakly supervised detection and segmentation, keyword cluster “weakly supervised · supervised semantic · image level”
State History Benchmarks Tried Baselines 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.
24 papers in 2023–24 and 7 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 20.9% to 9.7%.
Open questions the data points to
- Papers report Bilen et al. on PASCAL VOC 2007 test (mAP) anywhere from 27.7 to 34.5: comparisons against it depend on whose number you take.
What looks settled
- DOTA-v1.0 (AP50): best up to 2024 80.61, best in 2025–26 71.44 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: PASCAL VOC 2007 test (mAP, last in 2021); PASCAL VOC 2007 trainval (CorLoc, last in 2021); PASCAL VOC 2012 test (mAP, last in 2021); PASCAL VOC 2012 trainval (CorLoc, last in 2021); DUT-OMRON (MAE, 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 · Track and Transfer: Watching Videos to Simulate Strong Human Supervision for Weakly-Supervised Object Detection (CVPR 2016 · 79 citations)
- 2017 · Learning to Detect Salient Objects With Image-Level Supervision (CVPR 2017 · 1,320 citations)
- 2018 · Cross-Domain Weakly-Supervised Object Detection Through Progressive Domain Adaptation (CVPR 2018 · 622 citations)
- 2019 · C-MIL: Continuation Multiple Instance Learning for Weakly Supervised Object Detection (CVPR 2019 · 263 citations)
- 2020 · Weakly-Supervised Salient Object Detection via Scribble Annotations (CVPR 2020 · 331 citations)
- 2021 · Open-Vocabulary Object Detection Using Captions (CVPR 2021 · 417 citations)
- 2022 · Self-Supervised Transformers for Unsupervised Object Discovery Using Normalized Cut (CVPR 2022 · 151 citations)
- 2023 · Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection With Single Point Supervision (CVPR 2023 · 135 citations)
- 2024 · PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · 50 citations)
- 2025 · RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · 19 citations)
- 2026 · Bootstrapping MLLM for Weakly‑Supervised Class‑Agnostic Object Counting (ICLR 2026 · 0 citations)
Earliest papers here
- Track and Transfer: Watching Videos to Simulate Strong Human Supervision for Weakly-Supervised Object Detection (CVPR 2016 · 79 citations)
- Weakly Supervised Deep Detection Networks (CVPR 2016 · 63 citations)
- Saliency Guided Dictionary Learning for Weakly-Supervised Image Parsing (CVPR 2016 · 50 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.
PASCAL VOC 2007 test · mAP (higher is better)
14 papers, 2016–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 50 | We Don't Need No Bounding-Boxes: Training Object Class Detectors Using Only Human Verification (CVPR 2016 · 18 citations) |
| 2017 | 57.5 | Training Object Class Detectors With Click Supervision (CVPR 2017 · 119 citations) |
| 2018 | 80.7 | Pseudo Mask Augmented Object Detection (CVPR 2018 · 52 citations) |
PASCAL VOC 2007 trainval · CorLoc (higher is better)
11 papers, 2017–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 78.5 | Training Object Class Detectors With Click Supervision (CVPR 2017 · 119 citations) |
PASCAL VOC 2012 test · mAP (higher is better)
10 papers, 2018–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 79.5 | Pseudo Mask Augmented Object Detection (CVPR 2018 · 52 citations) |
PASCAL VOC 2012 trainval · CorLoc (higher is better)
7 papers, 2018–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 69.4 | W2F: A Weakly-Supervised to Fully-Supervised Framework for Object Detection (CVPR 2018 · 144 citations) |
| 2019 | 70.2 | Dissimilarity Coefficient Based Weakly Supervised Object Detection (CVPR 2019 · 93 citations) |
| 2021 | 71.9 | GradingNet: Towards Providing Reliable Supervisions for Weakly Supervised Object Detection by Grading the Box Candidates (AAAI 2021 · 20 citations) |
DOTA-v1.0 · AP50 (higher is better)
5 papers, 2023–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 80.61 | H2RBox-v2: Incorporating Symmetry for Boosting Horizontal Box Supervised Oriented Object Detection (NeurIPS 2023 · 17 citations) |
DUT-OMRON · MAE (lower is better)
4 papers, 2018–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 0.1 | Weakly Supervised Salient Object Detection Using Image Labels (AAAI 2018 · 131 citations) |
| 2021 | 0.0602 | Structure-Consistent Weakly Supervised Salient Object Detection with Local Saliency Coherence (AAAI 2021 · 163 citations) |
What has been tried
95 papers here report 1,070 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, 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 |
|---|---|---|---|---|
| Add SSC loss | DIOR-R testing set · mAP50 | 28.44 -> 32.98 | Helped | PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · 50 citations) |
| Use one sketch pattern per category instead of rectangles and circles with curve textures | DOTA-v1.0 · AP50 | 34.07 -> 40.27 | Helped | Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · 43 citations) |
| Add LGFE | MSD · S-measure (Sm) | 0.793 -> 0.810 | Helped | Weakly-Supervised Mirror Detection via Scribble Annotations (AAAI 2024 · 21 citations) |
| add DAFE | PASCAL VOC 2007 · mAP | 63.5 -> 64.1 | Helped | Weakly Supervised Open-Vocabulary Object Detection (AAAI 2024 · 20 citations) |
| Add unit cycle loss with λuc=0.05 instead of removing it | RSAR · AP75 | 24.07 -> 26.17 | Helped | RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · 19 citations) |
| add Gaussian overlap loss and Voronoi watershed loss | DOTA-v1.0 · AP50 | 0.00 -> 52.98 | Helped | Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · 14 citations) |
| Replace ALN-generated pseudo boxes with random-crop boxes during pretraining | ImageNetLoc-FS 1-shot · AP50 | 45.3 -> 5.1 | Hurt | Weakly Supervised Few-Shot Object Detection with DETR (AAAI 2024 · 9 citations) |
| add M2B transformation | PCSOD test set · Fm | 0.614 -> 0.736 | Helped | WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · 8 citations) |
| Add LHST and LHPG to the detector | Object365 · AP50 | 26.5 -> 32.5 | Helped | Open-Vocabulary Object Detection via Language Hierarchy (NeurIPS 2024 · 7 citations) |
| use prompt-based 3D shapes instead of ModelNet40 shapes | ScanNet val · mAP@0.25 | 30.6 -> 31.1 | Helped | Prompt3D: Random Prompt Assisted Weakly-Supervised 3D Object Detection (CVPR 2024 · 5 citations) |
| Replace RGB view with noise view | CASIAv1 · I-F1 | 0.720 -> 0.634 | Hurt | M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection (AAAI 2025 · 4 citations) |
| Add ABBS to the method | DIOR-R · 3-AP50 | 55.23 -> 62.13 | Helped | ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · 4 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 |
|---|---|---|---|---|
| OICR | PASCAL VOC 2007 test | mAP | 37.9 – 41.2 | 3 papers, e.g. CVPR 2018 and CVPR 2018 |
| Bilen et al. | PASCAL VOC 2007 test | mAP | 27.7 – 34.5 | 2 papers, e.g. CVPR 2016 and CVPR 2017 |
| OICR | PASCAL VOC 2007 | mAP | 41.2 – 47.93 | 2 papers, e.g. CVPR 2019 and ICLR 2023 |
| Fully supervised | PASCAL VOC 2007 test | mAP | 51 – 55.5 | 2 papers, e.g. CVPR 2016 and CVPR 2017 |
What reviewers will expect
Datasets and baselines named in the 6 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- DOTA-v1.0 · 4 of 6 papers
- DIOR-R · 2 of 6 papers
- DIOR · 2 of 6 papers
- HRSC · 2 of 6 papers
- RSAR · 2 of 6 papers
Baselines
- H2RBox · 4 of 6 papers
- FCOS · 3 of 6 papers
- RetinaNet · 3 of 6 papers
- KLD · 2 of 6 papers
- GWD · 2 of 6 papers
- KFIoU · 2 of 6 papers
- H2RBox-v2 · 2 of 6 papers
- PointOBB · 2 of 6 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 31 papers; from OpenAlex and the papers' affiliations).
- Harbin Institute of Technology · 5 papers
- Shanghai Jiao Tong University · 4 papers
- Shanghai AI Laboratory · 3 papers
- Southeast University · 3 papers
- Nanyang Technological University · 2 papers
- Beihang University · 2 papers
- City University of Hong Kong · 2 papers
- Tongji University · 2 papers
- East China Normal University · 2 papers
- Baidu · 2 papers
Use it
Download BibTeX of 37 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
- Weakly supervised semantic segmentation (116 papers)
- Weakly supervised object localization (54 papers)
- CAM refinement and prototypes (52 papers)
Elsewhere in the topic
- Pseudo-labeling for semi-supervised detection (76 papers)
- Consistency regularization for segmentation (25 papers)
- Semi-supervised instance segmentation (16 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).