CLEVR

CLEVR Dataset Overview

Dataset Description

• Name: CLEVR Dataset
• Purpose: Diagnostic for compositional language and elementary visual reasoning
• Origin: Proposed by Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Fei‑Fei Li, C Lawrence Zitnick, Ross Girshick at CVPR 2017

Dataset Generation

• Image Generation: Images rendered with Blender; a JSON file containing scene information for each image is provided.
• Question Generation: Questions, functional programs, and answers are generated from the scene information; a JSON file containing all questions is provided.

Dataset Content Examples

• Image Examples: Several synthetic images such as images/img1.png … images/img6.png.
• Question & Answer Examples:
  • Q: How many small spheres are there?
  • A: 2
  • Q: How many cubes are small objects or red metallic objects?
  • A: 2
  • Q: Do the metal sphere and the metal cylinder share the same color?
  • A: Yes
  • Q: Are there more small cylinders than metal objects?
  • A: No
  • Q: Is there a shiny cube to the right of the blue ball behind the large yellow object?
  • A: Yes

Citation

@inproceedings{johnson2017clevr,
  title={CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning},
  author={Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens
          and Fei‑Fei, Li and Zitnick, C Lawrence and Girshick, Ross},
  booktitle={CVPR},
  year={2017}
}