Research map / Vision-language and multimodal / Visual question answering
A-OKVQA: A Benchmark for Visual Question Answering Using World Knowledge
Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, Roozbeh Mottaghi
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 ground-truth rationales as input (ablation) | GPV-2 · A-OKVQA multiple choice (test) | accuracy: 53.7 -> 67.2 | Improved |
| add ground-truth rationales as input (ablation) | GPV-2 · A-OKVQA direct answer (test) | accuracy: 40.7 -> 51.7 | Improved |
| mask the answer inside ground-truth rationales (vs unmasked) (ablation) | GPV-2 + GT Ratl. · A-OKVQA multiple choice (test) | accuracy: 67.2 -> 58.3 | Hurt |
| mask answer in rationales vs no rationales (ablation) | GPV-2 · A-OKVQA multiple choice (test) | accuracy: 53.7 -> 58.3 | Improved |
| replace GT captions with GT rationales in the GPT-3 prompt (ablation) | GPT-3 · A-OKVQA multiple choice (val) | accuracy: 45.40 -> 56.74 | Improved |
| replace generated captions with generated rationales in the GPT-3 prompt (ablation) | ClipCap -> GPT-3 · A-OKVQA multiple choice (test) | accuracy: 43.61 -> 43.84 | No clear effect |
| add image input to question-only CLIP classifier (ablation) | CLIP classifier · A-OKVQA direct answer (test) | accuracy: 10.24 -> 14.27 | Improved |
| fine-tune ViLBERT on A-OKVQA after VQA training vs on A-OKVQA alone, direct answer (ablation) | ViLBERT · A-OKVQA direct answer (test) | accuracy: 25.9 -> 12.0 | Hurt |
| train CLIP contrastive MLP instead of zero-shot, image only | CLIP image-only · A-OKVQA multiple choice (test) | accuracy: 53.94 -> 50.09 | Hurt |
| add GPV-2 to the majority-vote ensemble | majority vote · A-OKVQA direct answer (test) | accuracy: 35.60 -> 40.7 | Hurt |
| majority vote of five weaker models vs best single of them | majority vote · A-OKVQA direct answer (test) | accuracy: 27.1 -> 29.5 | Improved |
| oracle selection across six models vs best single model | model ensemble · A-OKVQA direct answer (test) | accuracy: 40.7 -> 56.87 | Improved |
Closest papers by meaning
- OK-VQA: A Visual Question Answering Benchmark Requiring External Knowledge (CVPR 2019 · similarity 0.97)
- KVQA: Knowledge-Aware Visual Question Answering (AAAI 2019 · similarity 0.95)
- FVQA: Fact-Based Visual Question Answering (TPAMI 2017 · similarity 0.94)
- GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering (CVPR 2019 · similarity 0.94)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.94)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.93)
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.
