Contents Science Lab

Open Source: NU ICC: Image Collection Captioning Dataset

Abstract

We built the dataset upon the MS-COCO dataset by estimating the semantic contents of images and captions and using this to augment the dataset toward image collection captioning. To construct a dataset for the proposed task, we implement and compare two approaches based on image classification and image-caption retrieval.
• coco_resnet101_splited.json: is a dataset that is split by classifying the semantic contents of each image.
• coco_vsepp_splited.json: is a dataset that is split by estimating the semantic contents and captions with VSE++

Downloads

• Dataset
• License: CC BY-SA 4.0

Change Log

• Nov. 8, 2022: Web page opened

Citation

• Cite the following paper when you publish the paper of your research using this dataset.
  • Phueaksri Itthisak, Marc A. Kastner, Yasutomo Kawanishi, Takahiro Komamizu, Ichiro IDE: “Towards Captioning an Image Collection from a Combined Scene Graph Representation Approach”, In Proc: 29th International Conference on Multimedia Modeling, 2023.