Research map / Vision-language and multimodal / Visual question answering
Visual Question Answering as Reading Comprehension
Check 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 |
|---|---|---|---|
| Represent the image and supporting facts as text and answer with fine-tuned QANet | FVQA · FVQA test splits | top-1 accuracy (%): 62.20 -> 62.94 | Helped |
| Use SQuAD-pretrained QANet without FVQA fine-tuning instead of training from scratch on FVQA (ablation) | QANet · FVQA test splits | top-1 accuracy (%): 47.87 -> 55.14 | Helped |
| Fine-tune SQuAD-pretrained QANet on FVQA (ablation) | QANet · FVQA test splits | top-1 accuracy (%): 55.14 -> 62.94 | Helped |
| Use ground-truth region descriptions rather than predicted dense captions (ablation) | Open-ended VQA model · Visual Genome QA test | overall accuracy (%): 33.7 -> 44.8; 34.5 -> 44.8 | Helped |
| Use ground-truth region descriptions rather than predicted dense captions (ablation) | Multiple-choice VQA model · Visual7W Telling test | overall accuracy (%): 60.2 -> 69.8; 61.2 -> 69.8 | Helped |
| Use the better dense-caption model for predicted region descriptions (ablation) | Open-ended VQA model · Visual Genome QA test | overall accuracy (%): 33.7 -> 34.5 | Helped |
| Use the better dense-caption model for predicted region descriptions (ablation) | Multiple-choice VQA model · Visual7W Telling test | overall accuracy (%): 60.2 -> 61.2 | Helped |
| Use ground-truth region descriptions with the proposed reading-comprehension approach rather than VQA-Machine ground-truth facts | Open-ended VQA model · Visual Genome QA test | overall accuracy (%): 39.30 -> 44.8 | Helped |
| Use multiple-choice VQA with ground-truth descriptions | Visual7W Telling · Visual7W Telling test | overall accuracy (%): 67.1 -> 69.8 | Helped |
| Rely on predicted region descriptions instead of ground-truth descriptions (ablation) | Proposed VQA models · Visual Genome QA and Visual7W Telling | overall accuracy (%): Visual Genome: 44.8 vs. 33.7 and 34.5; Visual7W: 69.8 vs. 60.2 and 61.2 | Hurt |
| Use open-ended classification rather than selecting among provided candidate answers | VQA model · Visual Genome QA | answer handling | Hurt |
| Provide candidate answers as model inputs and encode question-answer interactions with context | Multiple-choice VQA model · Visual7W Telling | overall accuracy (%): 67.1 -> 69.8 | Helped |
Closest papers by meaning
- Answer Them All! Toward Universal Visual Question Answering Models (CVPR 2019 · similarity 0.94)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.93)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.92)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.92)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.91)
- Object Attribute Matters in Visual Question Answering (AAAI 2024 · similarity 0.91)
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.