Research map / Vision-language and multimodal / Human and embodied modeling
Describing Like Humans: On Diversity in Image Captioning
Qingzhong Wang, Antoni B. Chan
Check 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 |
|---|---|---|---|
| Use Self-CIDEr rather than mBLEU-mix to measure caption-set diversity | Caption diversity evaluation · Human diversity judgment | Pearson correlation: 0.585 -> 0.616 | Improved |
| Use Self-CIDEr rather than mBLEU-mix to measure caption-set diversity | Caption diversity evaluation · Human diversity judgment | Spearman correlation: 0.575 -> 0.617 | Improved |
| Use LSA rather than Self-CIDEr to measure caption-set diversity | Caption diversity evaluation · Human diversity judgment | Average per-image Spearman correlation: 0.674 -> 0.678 | Improved |
| Balance cross-entropy loss with CIDEr reward during RL training | Att2in and FC captioners · MSCOCO Karpathy test split | Diversity and accuracy tradeoff | Improved |
| Increase the CIDEr reward weight (ablation) | Att2in and FC captioners · MSCOCO Karpathy test split | Diversity | Hurt |
| Increase the CIDEr reward weight (ablation) | Att2in and FC captioners · MSCOCO Karpathy test split | Accuracy | Improved |
| Add retrieval reward to the combined loss (ablation) | Att2in and FC captioners · MSCOCO Karpathy test split | Diversity and accuracy | Improved |
| Use a very large retrieval-reward weight (ablation) | Att2in and FC captioners · MSCOCO Karpathy test split | Caption quality | Hurt |
| Use random sampling instead of a deterministic single caption (ablation) | Captioning models · MSCOCO Karpathy test split | Diversity | Improved |
| Use random sampling instead of a deterministic single caption (ablation) | Captioning models · MSCOCO Karpathy test split | Accuracy | Hurt |
| Use randomly cropped images to generate captions (ablation) | Captioning models · MSCOCO Karpathy test split | Accuracy | Improved |
| Use retrieval reward rather than CIDEr reward (ablation) | Att2in and FC captioners · MSCOCO Karpathy test split | Diversity | Improved |
Closest papers by meaning
- On Diversity in Image Captioning: Metrics and Methods (TPAMI 2020 · similarity 0.95)
- Compare and Reweight: Distinctive Image Captioning Using Similar Images Sets (ECCV 2020 · similarity 0.92)
- Semantic and Expressive Variations in Image Captions Across Languages (CVPR 2025 · similarity 0.87)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.87)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.86)
- CapsFusion: Rethinking Image-Text Data at Scale (CVPR 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.
