Research map / Vision-language and multimodal / Visual question answering
Dynamic Capsule Attention for Visual Question Answering
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 |
|---|---|---|---|
| replace stacked 2-layer attention (SAN) with single-layer dynamic capsule attention | VQA model · COCO-QA Color | accuracy: 57.9 -> 61.3 | Helped |
| replace stacked 2-layer attention (SAN) with single-layer dynamic capsule attention | VQA model · COCO-QA Objects | accuracy: 64.5 -> 65.0 | Helped |
| increase routing iterations from 1 to 3 (ablation) | CapsAtt · COCO-QA Color | accuracy: 57.0 -> 61.3 | Helped |
| increase routing iterations from 3 to 4 (ablation) | CapsAtt · COCO-QA | accuracy: Color 61.3 -> 62.0; Objects 65.0 -> 64.8 | Mixed |
| replace stacked 2-layer attention (SAN) with CapsAtt-Iter3 | VQA model · VQA2.0 test-dev | accuracy (All): 58.9 -> 62.1 | Helped |
| replace stacked 2-layer attention (SAN) with CapsAtt-Iter3 | VQA model · VQA1.0 test-dev | accuracy (All): 58.7 -> 63.9 | Helped |
| use Faster R-CNN regional features instead of ResNet grid features | CapsAtt-Iter3 · VQA1.0 test-dev | accuracy (All): 63.9 -> 66.1 | Helped |
| use Faster R-CNN regional features instead of ResNet grid features | CapsAtt-Iter3 · VQA2.0 test-dev | accuracy (All): 62.1 -> 65.5 | Helped |
| increase hidden dimensions from default (10.8M) to 25.6M params (ablation) | CapsAtt-Iter3 · VQA1.0 test-dev | accuracy: 66.10 -> 66.18 | No clear effect |
| shrink dimensions to 4.2M params (ablation) | CapsAtt-Iter3 · VQA1.0 test-dev | accuracy: 66.10 -> 65.22 | Hurt |
| shrink dimensions to 3.4M params (ablation) | CapsAtt-Iter3 · VQA1.0 test-dev | accuracy: 66.10 -> 64.28 | Hurt |
| use CapsAtt-Iter2 instead of single-layer Show-Attend-Tell attention with FRCNN features | image captioning · COCO Captions | CIDEr: 107 -> 108 | Helped |
Closest papers by meaning
- Multi-Level Attention Networks for Visual Question Answering (CVPR 2017 · similarity 0.91)
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.91)
- Deep Modular Co-Attention Networks for Visual Question Answering (CVPR 2019 · similarity 0.91)
- Tips and Tricks for Visual Question Answering: Learnings From the 2017 Challenge (CVPR 2018 · similarity 0.91)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.90)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.90)
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.