Research map / Vision-language and multimodal / Human and embodied modeling
Controllable Contextualized Image Captioning: Directing the Visual Narrative through User-Defined Highlights
Shunqi Mao, Chaoyi Zhang, Hang Su, Hwanjun Song, Igor Shalyminov, Weidong Cai
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 highlight-aware controllers (P-Ctrl or R-Ctrl) instead of extracting highlight-centered context | LongT5-base captioner · Wiki-Web2M full, single highlight | highlight recall: 18.45 -> 57.07 (P-Ctrl), 64.88 (R-Ctrl) | Improved |
| truncate context to highlight-centered sentences | LongT5 CIC model · Wiki-Web2M full | highlight control and caption quality: R 17.62-18.45 vs 57.07 (P-Ctrl) | Hurt |
| use R-Ctrl instead of P-Ctrl | Ctrl-CIC controllers · Wiki-Web2M full, single highlight | highlight recall vs Div-2: R 57.07 -> 64.88; Div-2 10.54 -> 10.30 | Mixed |
| fine-tune controllers for highlights instead of zero-shot LLaVA-1.5 | Ctrl-CIC · Wiki-Web2M full, single highlight | CLIPScore: 65.56 -> 60.91 (P-Ctrl) | Hurt |
| use R-Ctrl instead of CIC ground truth caption | Ctrl-CIC · GPT-4V evaluation | Context Relevance: 1 -> 1.01 | No clear effect |
| use Ctrl-CIC controllers on standard CIC task instead of LongT5 finetune | P-Ctrl · Wiki-Web2M test CIC | BLEU-4: 10.31 -> 10.88 | No clear effect |
| use P-Ctrl on standard CIC task instead of LongT5 finetune | P-Ctrl · Wiki-Web2M test CIC | CIDEr: 142.77 -> 130.80 | Hurt |
| remove visual input (text-only Llama-2) (ablation) | Ctrl-CIC · Wiki-Web2M full | Image Consistency (GPT-4V): 1 -> 0.61 | Hurt |
Closest papers by meaning
- Human-Like Controllable Image Captioning With Verb-Specific Semantic Roles (CVPR 2021 · similarity 0.90)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models (ICML 2025 · similarity 0.86)
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.86)
- Patch Matters: Training-free Fine-grained Image Caption Enhancement via Local Perception (CVPR 2025 · similarity 0.86)
- CALVIN: Improved Contextual Video Captioning via Instruction Tuning (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.
