CLEVR-DC

CLEVR-DC dataset used for paper:

Viewpoint-Agnostic Change Captioning with Cycle Consistency
Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, Gunhee Kim
To appear at ICCV 2021

Reference

If you find this repository useful, please cite the following paper:

@inproceedings{kim2021viewpoint,
  title={Viewpoint-Agnostic Change Captioning with Cycle Consistency},
  author={Kim, Hoeseong and Kim, Jongseok and Lee, Hyungseok and Park, Hyunsung
          and Kim, Gunhee}
  booktitle={ICCV},
  year={2021}
}

Introduction

CLEVR-DC is a CLEVR dataset for change captioning under drastic viewpoint changes. In contrast to other datasets with relatively small camera jitters, we reposition the camera to a random location in the after image. For the after scene, we perform one of the following:

• Change the color of one of the objects
• Change the texture (material) of one of the objects
• Add a random object
• Remove a random object
• Move a random object
• Do nothing (distractor)

We generate 8,000 images for each action. The split we used is included in split.json.

Examples