Dataset Card for CompreCap

Dataset Description

The CompreCap benchmark is characterized by human-annotated scene graph and focuses on the evaluation of comprehensive image captioning. It provides new semantic segmentation annotations for common objects in images, with an average mask coverage of 95.83%. Beyond the careful annotation of objects, CompreCap also includes high-quality descriptions of the attributes bound to the objects, as well as directional relation descriptions between the objects, composing a complete and directed scene graph structure:

The annotations of segmentation masks, category names, the descriptions of attributes and relationships are saved in ./anno.json. Based on the CompreCap benchmark, researchers can comprehensively accessing the quality of image captions generated by large vision-language models. The evaluation code is available here.

Licensing Information

We distribute the image under a standard Creative Common CC-BY-4.0 license. The individual images are under their own copyrights.

Citation

BibTeX:

@article{CompreCap,
  title={Benchmarking Large Vision-Language Models via Directed Scene Graph for Comprehensive Image Captioning},
  author={Fan Lu, Wei Wu, Kecheng Zheng, Shuailei Ma, Biao Gong, Jiawei Liu, Wei Zhai, Yang Cao, Yujun Shen, Zheng-Jun Zha},
  booktitle={arXiv},
  year={2024}
}