Research map / Graph learning / Knowledge and temporal graphs
Scene graph generation
Generate scene graphs with unbiased relation learning and transformers.
70 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Knowledge and temporal graphs, keyword cluster “scene · image · visual”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
17 papers in 2023–24 and 18 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 6.6% to 6.4%.
No longer reported by papers here: Visual Genome PredCls (mR@100, last in 2023); Visual Genome SGCls (mR@100, last in 2023); Visual Genome SGDet (mR@100, last in 2023); Visual Genome PredCls (mR@50, last in 2022); Visual Genome PredCls (R@100, last in 2022); Visual Genome SGCls (R@100, 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 · Scene Labeling Using Sparse Precision Matrix (CVPR 2016 · 19 citations)
- 2017 · Scene Graph Generation by Iterative Message Passing (CVPR 2017 · 1,254 citations)
- 2018 · Neural Motifs: Scene Graph Parsing With Global Context (CVPR 2018 · 1,124 citations)
- 2019 · Auto-Encoding Scene Graphs for Image Captioning (CVPR 2019 · 860 citations)
- 2020 · Unbiased Scene Graph Generation From Biased Training (CVPR 2020 · 755 citations)
- 2021 · Bipartite Graph Network With Adaptive Message Passing for Unbiased Scene Graph Generation (CVPR 2021 · 253 citations)
- 2022 · SGTR: End-to-End Scene Graph Generation With Transformer (CVPR 2022 · 124 citations)
- 2023 · Devil's on the Edges: Selective Quad Attention for Scene Graph Generation (CVPR 2023 · 52 citations)
- 2024 · Panoptic Scene Graph Generation with Semantics-Prototype Learning (AAAI 2024 · 29 citations)
- 2025 · Asynchronous Collaborative Graph Representation for Frames and Events (CVPR 2025 · 5 citations)
- 2026 · Panoptic Pairwise Distortion Graph (ICLR 2026 · 0 citations)
Earliest papers here
- Scene Labeling Using Sparse Precision Matrix (CVPR 2016 · 19 citations)
- Scene Graph Generation by Iterative Message Passing (CVPR 2017 · 1,254 citations)
- Neural Motifs: Scene Graph Parsing With Global Context (CVPR 2018 · 1,124 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.
Visual Genome PredCls · mR@100 (higher is better)
9 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 29.1 | Unbiased Scene Graph Generation From Biased Training (CVPR 2020 · 755 citations) |
| 2021 | 32.9 | Bipartite Graph Network With Adaptive Message Passing for Unbiased Scene Graph Generation (CVPR 2021 · 253 citations) |
| 2022 | 40.3 | Fine-Grained Predicates Learning for Scene Graph Generation (CVPR 2022 · 52 citations) |
Visual Genome SGCls · mR@100 (higher is better)
9 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 14.9 | Unbiased Scene Graph Generation From Biased Training (CVPR 2020 · 755 citations) |
| 2021 | 20.45 | Energy-Based Learning for Scene Graph Generation (CVPR 2021 · 157 citations) |
| 2022 | 27.6 | Fine-Grained Predicates Learning for Scene Graph Generation (CVPR 2022 · 52 citations) |
Visual Genome SGDet · mR@100 (higher is better)
7 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 11.1 | Unbiased Scene Graph Generation From Biased Training (CVPR 2020 · 755 citations) |
| 2021 | 11.6 | Energy-Based Learning for Scene Graph Generation (CVPR 2021 · 157 citations) |
| 2022 | 20.3 | Fine-Grained Predicates Learning for Scene Graph Generation (CVPR 2022 · 52 citations) |
Visual Genome PredCls · mR@50 (higher is better)
6 papers, 2021–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 30.4 | Bipartite Graph Network With Adaptive Message Passing for Unbiased Scene Graph Generation (CVPR 2021 · 253 citations) |
| 2022 | 36.4 | Fine-Grained Predicates Learning for Scene Graph Generation (CVPR 2022 · 52 citations) |
Visual Genome PredCls · R@100 (higher is better)
5 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 53.08 | Scene Graph Generation by Iterative Message Passing (CVPR 2017 · 1,254 citations) |
| 2018 | 67.1 | Neural Motifs: Scene Graph Parsing With Global Context (CVPR 2018 · 1,124 citations) |
| 2021 | 71.7 | Joint Modeling of Visual Objects and Relations for Scene Graph Generation (NeurIPS 2021 · 5 citations) |
Visual Genome SGCls · R@100 (higher is better)
5 papers, 2017–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2017 | 24.38 | Scene Graph Generation by Iterative Message Passing (CVPR 2017 · 1,254 citations) |
| 2018 | 36.5 | Neural Motifs: Scene Graph Parsing With Global Context (CVPR 2018 · 1,124 citations) |
| 2021 | 44.2 | Joint Modeling of Visual Objects and Relations for Scene Graph Generation (NeurIPS 2021 · 5 citations) |
What has been tried
56 papers here report 599 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 6%, 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 |
|---|---|---|---|---|
| Add indistinguishable triplet transfer | PSG SGDet · mR@100 | 22.1 -> 29.2 | Helped | Panoptic Scene Graph Generation with Semantics-Prototype Learning (AAAI 2024 · 29 citations) |
| Add adaptive refinement without hierarchical prediction | Visual Genome, PredCls · mR@100 UC/C | 66.3 / 39.2 -> 67.1 / 39.8 | Helped | HiKER-SGG: Hierarchical Knowledge Enhanced Robust Scene Graph Generation (CVPR 2024 · 21 citations) |
| Add relation rescoring | Visual Genome test, SGDet · R@50 | 6.9 -> 25.9 | Helped | DSGG: Dense Relation Transformer for an End-to-end Scene Graph Generation (CVPR 2024 · 14 citations) |
| add event streams | DSEC-Detection · mAP | 0.266 -> 0.284 | Helped | Asynchronous Collaborative Graph Representation for Frames and Events (CVPR 2025 · 5 citations) |
| Replace KWGNN with GNN | VG SGDET · Recall@100 | 39.3 -> 30.2 | Hurt | Kumaraswamy Wavelet for Heterophilic Scene Graph Generation (AAAI 2024 · 4 citations) |
| Add MLLM-based plausibility post-processing to conformal triplet sets | VG150 SGDET · AvgSize (triplet set) | roughly halves set size (e.g., BGNN 971.69 to 464.11) | Helped | Conformal Prediction and MLLM aided Uncertainty Quantification in Scene Graph Generation (CVPR 2025 · 3 citations) |
| Remove EMA and use fixed class-specific thresholds | Visual Genome SGCls · mR@50 / 100 | 16.9 / 18.0 -> 6.0 / 7.7 | Hurt | Adaptive Self-training Framework for Fine-grained Scene Graph Generation (ICLR 2024 · 2 citations) |
| remove the context checker and randomly combine triplet labels in prompts | S1 and S2 · evaluation metrics | approximately 2% decrease | Hurt | Adaptive Visual Scene Understanding: Incremental Scene Graph Generation (NeurIPS 2024 · 2 citations) |
| Add learnable parameters to TF-l-IDF | Visual Genome PredCls · mR@50 | 31.6 -> 33.6 | Helped | Scene Graph Generation Strategy with Co-occurrence Knowledge and Learnable Term Frequency (ICML 2024 · 1 citation) |
| Add image-scene alignment loss during denoising | COCO-Stuff and Visual Genome · FID and IS | Helped | Scene Graph-Grounded Image Generation (AAAI 2025 · 0 citations) | |
| Add geometric spatial proximity matrix to lane aggregation | OpenLane-V2 subsetA · TOPll | 21.6 -> 24.1 | Helped | T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 2025 · 0 citations) |
| add Hybrid Relation Assignment | VG150 test · mR@50 | 12.9 -> 14.9 | Helped | Hydra-SGG: Hybrid Relation Assignment for One-stage Scene Graph Generation (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 8 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Visual Genome · 2 of 8 papers
- Action Genome · 2 of 8 papers
Baselines
- MOTIFS · 2 of 8 papers
- BGNN · 2 of 8 papers
- EGTR · 2 of 8 papers
- STTran · 2 of 8 papers
- DSGDetr · 2 of 8 papers
- TEMPURA · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 35 papers; from OpenAlex and the papers' affiliations).
- Korea Advanced Institute of Science and Technology · 2 papers
- Electronics and Telecommunications Research Institute · 2 papers
- National University of Singapore · 2 papers
- ShanghaiTech University · 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
- Graph reasoning for visual recognition (96 papers)
- Graph-based dialog systems (94 papers)
- Heterogeneous graph learning (66 papers)
- Social and behavior graphs (87 papers)
Elsewhere in the topic
- Continuous-time dynamic graph learning (132 papers)
- Knowledge graph embeddings (110 papers)
- Temporal knowledge graph reasoning (84 papers)
- Temporal graph neural architectures (69 papers)
- Logical query answering and rule reasoning (68 papers)
- Spatio-temporal graph traffic forecasting (60 papers)
- Inductive knowledge graph completion (38 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).