Research map / Vision-language and multimodal / Human and embodied modeling
Context-Aware Visual Policy Network for Fine-Grained Image Captioning
Zheng-Jun Zha, Daqing Liu, Hanwang Zhang, Yongdong Zhang, Feng He Wu
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 explicit visual context via CAVP sub-policies (vs single sub-policy) (ablation) | Up-Down · MS-COCO Karpathy test (ablation) | CIDEr: 121.9 -> 126.3 | Improved |
| memorize full historical visual context rather than only last visual feature (ablation) | CAVP · MS-COCO Karpathy test (ablation) | CIDEr: 124.5 -> 126.3 | Improved |
| behavior cloning of output sub-policy from language-parser expert (ablation) | CAVP · MS-COCO Karpathy test (ablation) | CIDEr: 124.5 -> 124.6 | No clear effect |
| share LSTM weights among sub-policy networks (ablation) | CAVP · MS-COCO Karpathy test (ablation) | CIDEr: 126.4 -> 126.3 | No clear effect |
| share LSTM weights among sub-policy networks (ablation) | CAVP · Efficiency | parameters (M): 108.5M -> 83.0M | Improved |
| add CAVP to Up-Down | Up-Down · Efficiency | testing time (ms/image): 44.40 -> 56.48 | Hurt |
| optimize CIDEr instead of BLEU reward (ablation) | CAVP · MS-COCO Karpathy test (ablation) | CIDEr: 114.5 -> 126.4 | Improved |
| optimize SPIDEr instead of CIDEr reward (ablation) | CAVP · MS-COCO Karpathy test (ablation) | SPICE: 21.6 -> 23.1 | Improved |
| optimize METEOR reward instead of CIDEr (ablation) | CAVP · MS-COCO Karpathy test (ablation) | BLEU-4: 38.3 -> 33.6 | Hurt |
| add sentence-level CAVP visual context (ablation) | Single Policy · Stanford image-paragraph | CIDEr: 16.36 -> 17.29 | Improved |
| add hierarchical sentence-level and word-level visual policies (ablation) | Sentence CAVP · Stanford image-paragraph | CIDEr: 17.29 -> 19.63 | Improved |
| add CIDEr reinforcement training to hierarchical CAVP_XE (ablation) | Hierarchical CAVP · Stanford image-paragraph | CIDEr: 19.63 -> 20.94 | Improved |
Closest papers by meaning
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Fine-Grained Captioning of Long Videos through Scene Graph Consolidation (ICML 2025 · similarity 0.88)
- Benchmarking Large Vision-Language Models via Directed Scene Graph for Comprehensive Image Captioning (CVPR 2025 · similarity 0.88)
- Top-Down Semantic Refinement for Image Captioning (AAAI 2026 · similarity 0.88)
- FINECAPTION: Compositional Image Captioning Focusing on Wherever You Want at Any Granularity (CVPR 2025 · similarity 0.87)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · 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.
