Research map / Graph learning / Knowledge and temporal graphs
Pair Then Relation: Pair-Net for Panoptic Scene Graph Generation
Jinghao Wang, Zhengyu Wen, Xiangtai Li, Zujin Guo, Jingkang Yang, Ziwei Liu
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 |
|---|---|---|---|
| use Pair-Net pair-then-relation design instead of PSGFormer+ | PSG · PSG val | R@20: 18.9 -> 29.6 | Improved |
| use Pair-Net instead of PSGFormer+ | PSG · PSG val | Pair R@20: 28.6 -> 52.7 | Improved |
| use Swin-B backbone instead of ResNet-50 | Pair-Net · PSG val | R@20: 29.6 -> 33.3 | Improved |
| add BCE supervision of the pair matrix (ablation) | Pair Proposal Network · PSG val | mR@20: 0.5 -> 24.7 | Improved |
| add Matrix Learner (ablation) | Pair Proposal Network · PSG val | mR@20: 14.8 -> 24.7 | Improved |
| add linear embedding for subject and object (ablation) | Pair Proposal Network · PSG val | mR@20: 14.6 -> 24.7 | Improved |
| use CNN-tiny matrix learner instead of MLP (ablation) | Matrix Learner · PSG val | mR@20: 13.0 -> 24.7 | Improved |
| scale matrix learner from CNN-tiny to CNN-base (0.2M to 30M) (ablation) | Matrix Learner · PSG val | mR@20 / R@20: 24.7 / 29.6 -> 23.3 / 33.3 | Mixed |
| use Seesaw loss instead of cross entropy (ablation) | relation classification · PSG val | mR@20: 15.4 -> 24.7 | Improved |
| use concat subject-object pairs instead of image features as KV (ablation) | Relation Fusion · PSG val | mR@20: 1.3 -> 24.7 | Improved |
| change number of relation queries from 100 to 200 (ablation) | Relation Fusion · PSG val | mR@20: 24.7 -> 22.3 | No clear effect |
| increase pair loss weight from 5 to 10 (ablation) | Pair-Net · PSG val | mR@20: 24.7 -> 22.0 | No clear effect |
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.
- SGTR — SGTR: End-to-End Scene Graph Generation With Transformer (CVPR 2022) · better on 2 of 2 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
- Panoptic Scene Graph Generation (ECCV 2022 · similarity 0.93)
- OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models (ECCV 2024 · similarity 0.92)
- HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation (ICCV 2023 · similarity 0.92)
- DSGG: Dense Relation Transformer for an End-to-end Scene Graph Generation (CVPR 2024 · similarity 0.90)
- Panoptic Scene Graph Generation with Semantics-Prototype Learning (AAAI 2024 · similarity 0.89)
- Leveraging Predicate and Triplet Learning for 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.
