Research map / Graph learning / Knowledge and temporal graphs
Revisiting Scene Graph Generation from the Perspective of Detector-Conditioned Reachability
Runfeng Qu, Pia Bideau, Ole Hall, Julie Ouerfelli-Ethier, Klaus Obermayer, Olaf Hellwich
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 bottom-up queries to top-down-only model (ablation) | Dual-SGG · Visual Genome test | R@50/100: 31.9/36.3 -> 33.5/38.5 | Improved |
| add top-down queries to bottom-up-only model (ablation) | Dual-SGG · Visual Genome test | R@50/100: 31.3/34.7 -> 33.5/38.5 | Improved |
| add bottom-up queries (ablation) | Dual-SGG · Visual Genome test | micro-UDR@50/100: 1.5/2.2 -> 3.7/5.2 | Improved |
| add top-down queries (ablation) | Dual-SGG · Visual Genome test | micro-DR@50/100: 44.1/49.0 -> 47.3/54.3 | Improved |
| add dual-query design to detector-based baseline | Dual-SGG · Visual Genome test | micro-UDR@50/100: 0.2/0.3 -> 3.7/5.2 | Improved |
| add bottom-up queries under stronger Swin-B detector (ablation) | Dual-SGG* · Visual Genome test | R@50/100: 37.1/42.7 -> 38.8/44.6 | Improved |
| replace post-hoc max fusion with joint dual-query decoding (ablation) | Dual-SGG · Visual Genome test | R@50/100: 31.7/36.3 -> 33.5/38.5 | Improved |
| replace post-hoc average fusion with joint dual-query decoding (ablation) | Dual-SGG · Visual Genome test | R@50/100: 17.6/29.7 -> 33.5/38.5 | Improved |
| replace post-hoc normalization fusion with joint dual-query decoding (ablation) | Dual-SGG · Visual Genome test | R@50/100: 32.2/37.3 -> 33.5/38.5 | Improved |
| add content score S^ct to EPS (ablation) | Dual-SGG · Visual Genome test | R@50: 33.0 -> 33.5 | Improved |
| add spatial score S^sp to EPS (ablation) | Dual-SGG · Visual Genome test | R@50: 31.5 -> 33.5 | Improved |
| add self-attention mask blocking TD-to-BU information flow (ablation) | Dual-SGG · Visual Genome test | R@50: 32.8 -> 33.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Hydra-SGG — Hydra-SGG: Hybrid Relation Assignment for One-stage Scene Graph Generation (ICLR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Can We Build Scene Graphs, Not Classify Them? FlowSG: Progressive Image-Conditioned Scene Graph Generation with Flow Matching (CVPR 2026 · similarity 0.91)
- Adaptive Visual Scene Understanding: Incremental Scene Graph Generation (NeurIPS 2024 · similarity 0.89)
- Iterative Scene Graph Generation (NeurIPS 2022 · similarity 0.89)
- HiKER-SGG: Hierarchical Knowledge Enhanced Robust Scene Graph Generation (CVPR 2024 · similarity 0.88)
- Conformal Prediction and MLLM aided Uncertainty Quantification in Scene Graph Generation (CVPR 2025 · similarity 0.88)
- DSGG: Dense Relation Transformer for an End-to-end Scene Graph Generation (CVPR 2024 · similarity 0.87)
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.
