Research map / Vision-language and multimodal / Human and embodied modeling
Human Attention in Image Captioning: Dataset and Analysis
Sen He, Hamed R. Tavakoli, Ali Borji, Nicolas Pugeault
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 free-viewing saliency guidance (foveated fixation locations) to soft-attention captioning | soft-attention captioning, ResNet-18 · Flickr30k test | CIDEr: 0.352 -> 0.403 | Improved |
| add task saliency (trained on capgaze2 captioning fixations) to soft-attention captioning | soft-attention captioning, ResNet-18 · Flickr30k test | CIDEr: 0.352 -> 0.416 (+18.2%) | Improved |
| add task saliency (trained on capgaze2 captioning fixations) to soft-attention captioning | soft-attention captioning, ResNet-18 · MSCOCO test | CIDEr: 0.81 -> 0.907 (+12%) | Improved |
| add free-viewing saliency guidance to soft-attention captioning | soft-attention captioning, ResNet-18 · MSCOCO test | CIDEr: 0.81 -> 0.889 | Improved |
| train the saliency model on captioning-task fixations (capgaze2) rather than free-viewing Salicon (ablation) | saliency-guided captioning · Flickr30k test | CIDEr: 0.403 -> 0.416 | Improved |
| train the saliency model on captioning-task fixations (capgaze2) rather than free-viewing Salicon (ablation) | saliency-guided captioning · MSCOCO test | CIDEr: 0.889 -> 0.907 | Improved |
| use a bottom-up saliency model (SalGAN) instead of top-down soft attention to predict captioning fixations | attention prediction · captioning fixations (NSS) | NSS: 1.128 -> 1.618 | Improved |
| compare attention under captioning task versus free-viewing for CNN-encoder agreement | VGG-16 conv5-3 · human-attended regions covered by CNN | percentage: 72.5% -> 78.1% | Improved |
| use full-duration captioning fixations instead of first 3 seconds for CNN agreement | VGG-16 conv5-3 · human-attended regions covered by CNN | percentage: 78.1% -> 77.9% | No clear effect |
| correlate machine-human spatial attention congruency with caption quality | soft-attention captioning · per-image CIDEr vs NSS | Spearman correlation: 0.01 | No clear effect |
| correlate machine-human temporal attention distance with caption quality | soft-attention captioning · per-image CIDEr vs DTW distance | Spearman correlation: -0.05 | No clear effect |
| describe the image while viewing (captioning) instead of free-viewing, over full duration | human attention · attention ratio on described objects | attention ratio: 0.66 -> 0.63 | Hurt |
Closest papers by meaning
- Boosted Attention: Leveraging Human Attention for Image Captioning (ECCV 2018 · similarity 0.94)
- Paying Attention to Descriptions Generated by Image Captioning Models (ICCV 2017 · similarity 0.92)
- Attention Alignment Between Humans and Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models (ICML 2025 · similarity 0.88)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.87)
- Semantic and Expressive Variations in Image Captions Across Languages (CVPR 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.
