hoveringgull/COCO_GridQA

COCO-GridQA Dataset

Overview

The COCO-GridQA dataset is a derived dataset created from the COCO (Common Objects in Context) validation set. It focuses on spatial reasoning tasks by arranging object crops from COCO images into a 2x2 grid and providing question-answer pairs about the positions of objects within the grid.

This dataset is designed for tasks such as spatial reasoning, visual question answering (VQA), and object localization. Each sample consists of:

• A composite image containing four object crops arranged in a 2x2 grid.
• Annotations that include questions like "Where is the tennis racket in the image?" and options like "Bottom-left", "Top-left", "Bottom-right" and "Top-right".

File Format

This dataset is provided in a TSV (Tab-Separated Values) format.

Column	Description
index	A unique identifier for each data sample
question	A textual question related to the content of the image
A	The first option in the multiple-choice answers
B	The second option in the multiple-choice answers
C	The third option in the multiple-choice answers
D	The fourth option in the multiple-choice answers
answer	The correct answer, represented by A, B, C, or D
image	A base64-encoded string representing the image

---

License

The COCO-GridQA dataset is derived from the COCO dataset, which is made available under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Key Points of the License:

• Attribution: You must give appropriate credit to the original COCO dataset creators whenever you use this dataset or any derivative works. Please cite the original COCO paper if you use this dataset in your research: