Research map / Graph learning / Knowledge and temporal graphs
In Defense of Scene Graphs for Image Captioning
Kien Nguyen, Subarna Tripathi, Bang Du, Tanaya Guha, Truong Q. Nguyen
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 bounding-box node grounding (ablation) | SG2Caps baseline · COCO Karpathy test (ablation, cross-entropy) | CIDEr: 92.6 -> 97.4 | Improved |
| add HOI graph on top of BBox (ablation) | SG2Caps · COCO Karpathy test (ablation, cross-entropy) | CIDEr: 97.4 -> 104.4 | Improved |
| add global visual feature summary node (ablation) | SG2Caps · COCO Karpathy test (ablation, cross-entropy) | CIDEr: 104.4 -> 106.6 | Improved |
| add global visual feature to baseline without bbox/HOI (ablation) | scene graph baseline · COCO Karpathy test (ablation, cross-entropy) | CIDEr: 92.6 -> 98.1 | Improved |
| add scene graph embeddings to global-feature model (ablation) | SG2Caps-G · COCO Karpathy test | CIDEr: 99.5 -> 106.6 | Improved |
| add HOI graph to pseudolabel input | TSG-trained caption model · COCO Karpathy test (TSG-trained model, different inputs) | CIDEr: 31.5 -> 38.6 | Improved |
| feed visual scene graph instead of textual scene graph at inference | TSG-trained caption model · COCO Karpathy test (TSG-trained model, different inputs) | CIDEr: 158.6 -> 31.5 | Hurt |
| train with CIDEr-based RL after cross-entropy | SG2Caps · COCO Karpathy test (scene graph only) | CIDEr: 104.4 -> 112.3 | Improved |
| use SG2Caps instead of SubGC in G-only mode | scene-graph-only captioner · COCO Karpathy test (scene graph only) | CIDEr: 98.9 -> 104.4 | Improved |
| use scene graph labels only instead of visual-feature model | SG2Caps vs SGAE · COCO Karpathy test | trainable parameters: 41M -> 21M | Improved |
Closest papers by meaning
- Comprehensive Image Captioning via Scene Graph Decomposition (ECCV 2020 · similarity 0.93)
- Unpaired Image Captioning via Scene Graph Alignments (ICCV 2019 · similarity 0.90)
- Benchmarking Large Vision-Language Models via Directed Scene Graph for Comprehensive Image Captioning (CVPR 2025 · similarity 0.89)
- SGDiff: Scene Graph Guided Diffusion Model for Image Collaborative SegCaptioning (AAAI 2025 · similarity 0.88)
- Fine-Grained Captioning of Long Videos through Scene Graph Consolidation (ICML 2025 · similarity 0.87)
- SCENIR: Visual Semantic Clarity through Unsupervised Scene Graph Retrieval (ICML 2025 · 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.
