Research map / Graph learning / Knowledge and temporal graphs
Generative Compositional Augmentations for Scene Graph Prediction
Boris Knyazev, Harm de Vries, Cătălina Cangea, Graham W. Taylor, Aaron Courville, Eugene Belilovsky
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GAN augmentation without perturbation (ablation) | IMP++ · Visual Genome all-shot SGCls | recall: 48.70 -> 50.38 | Improved |
| add GraphN perturbation (alpha=5) to GAN (ablation) | IMP++ GAN · Visual Genome zero-shot SGCls | recall: 9.25 -> 9.62 | Improved |
| add Rand perturbation to GAN (ablation) | IMP++ GAN · Visual Genome all-shot SGCls | recall: 50.38 -> 49.83 | Hurt |
| add Neigh perturbation to GAN (ablation) | IMP++ GAN · Visual Genome all-shot SGCls | recall: 50.38 -> 50.00 | Hurt |
| use GraphN instead of Rand perturbation (ablation) | IMP++ GAN · Visual Genome zero-shot PredCls | recall: 28.71 -> 29.18 | Improved |
| raise GraphN threshold alpha from 2 to 20 (ablation) | GAN+GraphN · Visual Genome zero-shot SGCls | recall: 9.89 -> 9.65 | Hurt |
| raise GraphN threshold alpha from 2 to 20 (ablation) | GAN+GraphN · Visual Genome 100-shot PredCls | recall: 69.17 -> 69.31 | Improved |
| add GAN+GraphN vs Reweight | IMP++ · Visual Genome all-shot PredCls | recall: 74.68 -> 78.94 | Improved |
| add TDE debiasing | IMP++ · Visual Genome zsR@100 PredCls | recall: 21.2 -> 21.5 | No clear effect |
| add GAN+GraphN | NM++ · Visual Genome zsR@100 PredCls | recall: 13.4 -> 17.4 | Improved |
| add GAN+GraphN | IMP++ · Visual Genome zsR@100 PredCls | recall: 21.2 -> 21.8 | Improved |
| increase top-k for Neigh perturbation from 2 to 20 (ablation) | GAN+Neigh · Visual Genome zero-shot SGCls | recall: 9.49 -> 9.65 | No clear effect |
Closest papers by meaning
- Generating Triples With Adversarial Networks for Scene Graph Construction (AAAI 2018 · similarity 0.93)
- Generative Scene Graph Networks (ICLR 2021 · similarity 0.91)
- Compositional GAN: Learning Image-Conditional Binary Composition (IJCV 2020 · similarity 0.90)
- Image Generation From Scene Graphs (CVPR 2018 · similarity 0.90)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.89)
- Scene Graph Disentanglement and Composition for Generalizable Complex Image Generation (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
