Research map / Vision-language and multimodal / Human and embodied modeling
Adaptively Aligned Image Captioning via Adaptive Attention Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 adaptive attention steps instead of one fixed step per word (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 123.76 ± 0.41 -> 126.48 ± 0.22 | Helped |
| Use adaptive attention steps instead of a fixed 4 steps per word (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 124.96 ± 0.33 -> 126.48 ± 0.22 | Helped |
| Use 8 fixed recurrent attention steps instead of 4 (ablation) | Image captioner · MS COCO Karpathy test split | CIDEr-D: 124.96 ± 0.33 -> 124.85 ± 0.30 | No clear effect |
| Increase the fixed recurrent attention steps from 2 to 4 (ablation) | Image captioner · MS COCO Karpathy test split | CIDEr-D: 124.22 ± 0.24 -> 124.96 ± 0.33 | Helped |
| Add attention time penalty with lambda 1e-4 rather than lambda 1e-1 (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 118.7 -> 126.7 | Helped |
| Reduce attention time penalty from lambda 1e-3 to 1e-4 (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 125.7 -> 126.7 | Helped |
| Remove the attention time penalty (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 126.7 -> 126.7 | No clear effect |
| Replace additive attention with dot-product attention using one head (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 126.7 -> 126.3 | Hurt |
| Increase dot-product attention heads from 1 to 8 (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 126.3 -> 128.6 | Helped |
| Increase dot-product attention heads from 8 to 16 (ablation) | AAT image captioner · MS COCO Karpathy test split | CIDEr-D: 128.6 -> 128.0 | Hurt |
| Use AAT rather than Up-Down | Image captioner · MS COCO Karpathy test split | CIDEr-D: 120.1 -> 128.6 | Helped |
| Use AAT rather than SGAE | Image captioner · MS COCO Karpathy test split | ROUGE-L: 58.6 -> 58.5 | Hurt |
Closest papers by meaning
- Attention-Aligned Transformer for Image Captioning (AAAI 2022 · similarity 0.89)
- Knowing When to Look: Adaptive Attention via a Visual Sentinel for Image Captioning (CVPR 2017 · similarity 0.89)
- Text-Guided Attention Model for Image Captioning (AAAI 2017 · similarity 0.87)
- RSTNet: Captioning With Adaptive Attention on Visual and Non-Visual Words (CVPR 2021 · similarity 0.87)
- Deliberate Attention Networks for Image Captioning (AAAI 2019 · similarity 0.87)
- Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models (ICML 2025 · similarity 0.85)
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.