Research map / Detection and segmentation / Open-set and few-shot detection
Large-Scale Visual Relationship Understanding
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 the proposed visual-semantic embedding model instead of KL distillation | VRD relationship detection · VRD, relationship detection, free k, R@50 | recall: 22.68 -> 26.98 | Helped |
| Use the proposed model instead of MotifNet-LeftRight | VG200 scene graph detection · VG200, scene graph detection, R@100 | recall: 30.3 -> 32.5 | Helped |
| Use the proposed model instead of 3-branch Fast-RCNN | VG80k all-class relationship recognition · VG80k, all classes, relationship triplet | top1 accuracy: 9.73 -> 15.72 | Helped |
| Use the proposed model instead of 3-branch Fast-RCNN | VG80k tail-class relationship recognition · VG80k, tail classes, relationship triplet | top5 accuracy: 3.24 -> 13.33 | Helped |
| Replace triplet loss with triplet-softmax loss (ablation) | VG80k relationship recognition · VG80k, all classes, relationship triplet | top1 accuracy: 8.01 -> 14.53 | Helped |
| Combine wiki embeddings with relationship-level co-occurrence embeddings (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top5 accuracy: 46.03 -> 48.83 | Helped |
| Add a learned semantic embedding layer rather than use raw word vectors (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top1 accuracy: 11.21 -> 15.72 | Helped |
| Use two semantic layers rather than one (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top5 accuracy: 48.23 -> 48.83 | Helped |
| Add both early and late visual feature concatenations (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top1 accuracy: 10.47 -> 15.72 | Helped |
| Add visual matching loss L_x to L_y alone (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top1 accuracy: 15.21 -> 15.07 | No clear effect |
| Add visual consistency loss L_c (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top1 accuracy: 15.07 -> 15.72 | Helped |
| Add visual matching loss L_x when L_y and L_c are present (ablation) | VG80k relationship recognition · VG80k, relationship triplet | top1 accuracy: 15.53 -> 15.72 | Helped |
Closest papers by meaning
- Detecting Visual Relationships With Deep Relational Networks (CVPR 2017 · similarity 0.91)
- Visual Relationship Detection With Deep Structural Ranking (AAAI 2018 · similarity 0.90)
- On Exploring Undetermined Relationships for Visual Relationship Detection (CVPR 2019 · similarity 0.90)
- Synthetic Visual Genome (CVPR 2025 · similarity 0.90)
- Towards Flexible Visual Relationship Segmentation (NeurIPS 2024 · similarity 0.89)
- Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (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.