Visual7W: Grounded Question Answering in Images

Yuke Zhu, Oliver Groth, Michael Bernstein and Li Fei-Fei

---

Data Collection Interfaces

Our data collection is conducted on Amazon Mechanical Turk (AMT) with simpleamt, an easy-to-use HIT management framework. Here we provide the data collection interfaces that we have used in Visual7W to collect human annotations from AMT workers. For more details about data collection, please check out our Visual Genome project.

QA Data Generation

• Freeform telling QA collection
• Telling QA based on Visual Genome regions
• Pointing QA based on Visual Genome regions
• Multiple choices for telling QA

QA Verification

• Verify telling QA quality
• Verify telling QA spell checking results (from Google search)
• Verify pointing QA quality
• Verify pointing QA spell checking results (from Google search)

Human Experiments

• Human experiments on telling QA without images
• Human experiments on telling QA with images
• Human experiments on pointing QA with images