Research map / Graph learning / Knowledge and temporal graphs
Iterative Scene Graph Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use iterative scene-graph refinement with cross-step and within-step conditioning plus joint loss (ablation) | Scene graph generation · Visual Genome | mR@20/50: 11.2 ~/~ 14.7 -> 11.8 ~/~ 15.8 | Helped |
| Add joint loss across decoder outputs and refinement layers (ablation) | Iterative scene graph model · Visual Genome | mR@20/50: 1.7 ~/~ 1.9 -> 11.2 ~/~ 14.7 | Helped |
| Add within-step conditioning (ablation) | Iterative scene graph model with joint loss · Visual Genome | mR@20/50: 11.2 ~/~ 14.7 -> 11.5 ~/~ 15.3 | Helped |
| Add conditioning across refinement steps (ablation) | Iterative scene graph model with joint loss · Visual Genome | mR@20/50: 11.2 ~/~ 14.7 -> 11.7 ~/~ 15.4 | Helped |
| Combine within-step and across-step conditioning (ablation) | Iterative scene graph model · Visual Genome | mR@20/50: 11.5 ~/~ 15.3 -> 11.8 ~/~ 15.8 | Helped |
| Add iterative refinement to MOTIF | MOTIF · Visual Genome | mR@20/50: 6.0 ~/~ 8.0 -> 6.4 ~/~ 8.5 | Helped |
| Continue refinement from step 1 to step 6 (ablation) | Ours (α=0.14, β=0.75) · Visual Genome test | hR@20/50: 12.4 ~/~ 16.9 -> 15.1 ~/~ 19.7 | Helped |
| Apply predicate loss re-weighting instead of α=0 (ablation) | Ours · Visual Genome test | mR@20/50: 5.8 ~/~ 8.0 -> 11.8 ~/~ 15.8 | Helped |
| Increase predicate loss weighting to α=0.21 (ablation) | Ours (β=0.75) · Visual Genome test | R@20/50: 21.0 ~/~ 26.1 -> 20.0 ~/~ 24.8 | Hurt |
| Use 150 or 450 decoder queries instead of 300 (ablation) | Ours (α=0.14, β=0.75) · Visual Genome test | hR@20/50: 15.1 ~/~ 19.7 -> 15.0 ~/~ 18.9; 14.9 ~/~ 19.6 | Hurt |
| Select the top 3 predicates per subject-object pair | Ours (α=0.14, β=0.75) · Visual Genome | R@50/100: 26.1 ~/~ 28.7 -> 30.8 ~/~ 35.6 | Helped |
| Use loss re-weighting with α=0.14, β=0.75 instead of α=0 (ablation) | Ours · Action Genome | mR@20: 10.9 -> 19.7 | Helped |
Closest papers by meaning
- IS-GGT: Iterative Scene Graph Generation With Generative Transformers (CVPR 2023 · similarity 0.92)
- A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)
- DSGG: Dense Relation Transformer for an End-to-end Scene Graph Generation (CVPR 2024 · similarity 0.90)
- Scene Graph-Grounded Image Generation (AAAI 2025 · similarity 0.89)
- R3CD: Scene Graph to Image Generation with Relation-Aware Compositional Contrastive Control Diffusion (AAAI 2024 · similarity 0.89)
- From Pixels to Graphs: Open-Vocabulary Scene Graph Generation with Vision-Language Models (CVPR 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.