[memo]A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge

Allen研究所, Dusting Schwenkが第一著者

Introduction

• Knowledge, common-sense in NLP. Creating benchmark which requires a diverse knowledge outside
• Detailed analysis
• An evaluation

BERT

CLIP

GPT-3

Related works

Knowledge-based VQA datasets

• FVQA does not require reasoning
• KVQA focuses too much on names of people.
• OK-VQA
• S3VQA requires detecting objects in images

Commonsense

• Wikipedia SQu-AD
• CommonsenseQA

Dataset collection

COCO2017

Interesting images are filtered by using Resnet

Clustered images by CLIP

Workers are hired

Experiment

BERT, CLIP, CLIPcap

Conclusion

Commonsenseとは．．．．となる