Research map / Vision-language and multimodal / Visual question answering
VrR-VG: Refocusing Visually-Relevant Relationships
Yuanzhi Liang, Yalong Bai, Wei Zhang, Xueming Qian, Li Zhu, Tao Mei
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 relation supervision (relation loss and relation data) to feature learning (ablation) | Ours features on VrR-VG, MUTAN · VQA-2.0 val | accuracy (All): 62.93 -> 64.57 | Improved |
| add relation supervision (relation loss and relation data) to feature learning (ablation) | Ours features on VrR-VG, MFH · VQA-2.0 val | accuracy (All): 64.68 -> 65.46 | Improved |
| filter visually-irrelevant relationships with VD-Net (R-VG to VrR-VG) (ablation) | Ours features, MUTAN · VQA-2.0 val | accuracy (All): 63.77 -> 64.57 | Improved |
| filter visually-irrelevant relationships with VD-Net (R-VG to VrR-VG) (ablation) | Ours features, MFH · VQA-2.0 val | accuracy (All): 64.22 -> 65.46 | Improved |
| train relation-aware features on VrR-VG instead of VG150 | Ours features, MUTAN · VQA-2.0 val | accuracy (All): 59.49 -> 64.57 | Improved |
| add relation supervision to feature learning (ablation) | Ours features on VrR-VG · MSCOCO 2014 captioning (cross-entropy loss) | CIDEr: 110.3 -> 114.0 | Improved |
| add relation supervision to feature learning (ablation) | Ours features on VrR-VG · MSCOCO 2014 captioning (CIDEr optimization) | CIDEr: 116.8 -> 120.7 | Improved |
| filter visually-irrelevant relationships with VD-Net (R-VG to VrR-VG) (ablation) | Ours features · MSCOCO 2014 captioning (cross-entropy loss) | CIDEr: 111.2 -> 114.0 | Improved |
| train relation-aware features on VrR-VG instead of VG150 | Ours features · MSCOCO 2014 captioning (CIDEr optimization) | CIDEr: 108.0 -> 120.7 | Improved |
| use relation-aware VrR-VG features instead of BottomUp object-attribute features on larger BottomUp-VG | MUTAN VQA · VQA-2.0 val | accuracy (All): 62.84 -> 64.57 | Improved |
| switch dataset from VG150 to VrR-VG for the frequency baseline | Frequency-Baseline · PredDet | R@50: 94.6 -> 69.8 | Hurt |
| switch dataset from VG150 to VrR-VG for a learnable method | Neural-Motifs · PredDet | R@50: 96.0 -> 87.6 | Hurt |
Closest papers by meaning
- Large-Scale Visual Relationship Understanding (AAAI 2019 · similarity 0.91)
- RelTransformer: A Transformer-Based Long-Tail Visual Relationship Recognition (CVPR 2022 · similarity 0.90)
- Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection (ECCV 2024 · similarity 0.89)
- Synthetic Visual Genome (CVPR 2025 · similarity 0.89)
- ART: Adaptive Relation Tuning for Generalized Relation Prediction (ICCV 2025 · similarity 0.89)
- Relation-aware Hierarchical Prompt for Open-vocabulary Scene Graph Generation (AAAI 2025 · 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.
