Research map / Detection and segmentation / Open-set and few-shot detection
Learning To Generate Language-Supervised and Open-Vocabulary Scene Graph Using Pre-Trained Visual-Semantic Space
Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, Chang-Wen Chen
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 VS3 with GLIP-L (Swin-L) instead of the strongest prior closed-set model | Fully supervised SGG · VG150 test, SGDET | R@100: 38.1 -> 41.5 | Improved |
| Transfer a pre-trained visual-semantic space to language-supervised SGG | VS3 · VG150 test, unlocalized graph supervision | R@100: 13.15 -> 28.19 | Improved |
| Transfer a pre-trained visual-semantic space to language-supervised SGG | VS3 · VG150 test, VG caption supervision | R@100: 12.14 -> 20.54 | Improved |
| Transfer a pre-trained visual-semantic space to language-supervised SGG | VS3 · VG150 test, COCO caption supervision | R@100: 7.62 -> 9.90 | Improved |
| Use a larger pre-trained backbone, Swin-L instead of Swin-T | VS3 · VG150 test, fully supervised SGDET | R@100: 39.2 -> 41.5 | Improved |
| Add visual region features to relation representations (ablation) | VS3 · VG150 test, fully supervised SGDET | R@100: 36.7 -> 39.2 | Improved |
| Add spatial features to relation representations (ablation) | VS3 · VG150 test, fully supervised SGDET | R@100: 37.8 -> 39.2 | Improved |
| Use frequency bias for predicate classification (ablation) | VS3(Swin-T) · VG150 test, language-supervised SGDET with unlocalized graphs | R@100: 28.19 -> 31.75 | Improved |
| Use all captions per image instead of one caption (ablation) | VS3 with simple parser · VG150 test, language-supervised SGDET on COCO captions | R@20/50/100: 5.07 / 6.25 / 7.36 -> 5.42 / 6.82 / 7.93 | Improved |
| Use the advanced parser instead of the simple parser (ablation) | VS3 trained on all COCO captions · VG150 test, language-supervised SGDET | R@20/50/100: 5.42 / 6.82 / 7.93 -> 5.59 / 7.30 / 8.62 | Improved |
| Use a larger pre-trained backbone, Swin-L instead of Swin-T | VS3 · VG150 Ov-SGG PREDCLS | R@50/100: 50.10 / 52.05 -> 55.88 / 58.18 | Improved |
| Train with VG captions rather than COCO captions | VS3(Swin-L) · VG150 language-supervised Ov-SGG | R@50/100: 6.76 / 8.45 -> 12.98 / 16.29 | Improved |
Closest papers by meaning
- From Pixels to Graphs: Open-Vocabulary Scene Graph Generation with Vision-Language Models (CVPR 2024 · similarity 0.93)
- Scene Graph Generation with Role-Playing Large Language Models (NeurIPS 2024 · similarity 0.92)
- Relation-aware Hierarchical Prompt for Open-vocabulary Scene Graph Generation (AAAI 2025 · similarity 0.91)
- RA-SGG: Retrieval-Augmented Scene Graph Generation Framework via Multi-Prototype Learning (AAAI 2025 · similarity 0.90)
- LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation (CVPR 2024 · similarity 0.90)
- Leveraging Predicate and Triplet Learning for Scene Graph Generation (CVPR 2024 · similarity 0.90)
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.
