Research map / Graph learning / Knowledge and temporal graphs
Topic Scene Graph Generation by Attention Distillation From Caption
Wenbin Wang, Ruiping Wang, Xilin 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 |
|---|---|---|---|
| add attention distillation from captions (weak supervision) to Up-Down captioner | UD-ICRC · important relationship discovery | mean recall: 12.64 -> 34.06 | Improved |
| add attention distillation from captions to Transformer captioner | Transformer-ICRC · important relationship discovery | mean recall: 9.82 -> 37.27 | Improved |
| use max pooling instead of mean pooling over time steps (ablation) | UD-ICRC-attn · important relationship discovery | mean recall: 12.84 -> 26.63 | Improved |
| mask non-noun words when collecting attention (ablation) | UD-ICRC-attn · important relationship discovery | mean recall: 34.06 -> 26.63 | Hurt |
| use subject, object and union features instead of union only (ablation) | UD-ICRC-attn · important relationship discovery | mean recall: 21.95 -> 33.76 | Improved |
| use subject, object and union features instead of subject and object only (ablation) | UD-ICRC-attn · important relationship discovery | mean recall: 25.98 -> 33.76 | Improved |
| add semantic embeddings of subject and object categories (ablation) | UD-ICRC-attn · important relationship discovery | mean recall: 33.76 -> 34.06 | Improved |
| add joint training with relational captions (lambda 0.1) to Up-Down captioner (ablation) | UD · image captioning | CIDEr-D: 94.1 -> 95.3 | Improved |
| add joint training with relational captions to Transformer captioner (ablation) | Transformer · image captioning | CIDEr-D: 85.6 -> 91.5 | Improved |
| increase relational caption loss weight lambda from 0.1 to 1.0 (ablation) | UD-ICRC · image captioning | CIDEr-D: 95.3 -> 93.5 | Hurt |
| increase relational caption loss weight lambda from 0.1 to 1.0 (ablation) | UD-ICRC · important relationship discovery | R@20: 3.45 -> 2.31 | Hurt |
| add joint training with image captions to Transformer relational captioner (ablation) | Transformer-RC · relational captioning | mAP: 5.26 -> 5.15 | No clear effect |
Closest papers by meaning
- Linguistic Structures As Weak Supervision for Visual Scene Graph Generation (CVPR 2021 · similarity 0.90)
- A Comprehensive Survey of Scene Graphs: Generation and Application (TPAMI 2021 · similarity 0.90)
- Sketching Image Gist: Human-Mimetic Hierarchical Scene Graph Generation (ECCV 2020 · similarity 0.89)
- Scene Graph Generation From Objects, Phrases and Region Captions (ICCV 2017 · similarity 0.88)
- A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models (NeurIPS 2026 · similarity 0.87)
- Scene Graph Generation with Role-Playing Large Language Models (NeurIPS 2024 · similarity 0.86)
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.
