xuemduan/reevaluate-image-text-pairs

REEVLAUATE Image-Text Pair Dataset

Overview

This is an image-text pair dataset constructed for the Knowledge-Enhanced Multimodal Retrieval System, built upon the REEVLAUATE KG ArtKB. The dataset is designed for training and evaluating the CLIP model for the retrieval system.

Data Source

The ArtKB knowledge base combines data from two primary sources:

• Wikidata
• Pilot Museums

Dataset Structure

The dataset is organized into three splits:

• Train: Training set
• Validation: Validation set
• Test: Test set

Each split contains:

• Images: Visual content stored in subdirectories (000/, 001/, ..., 999/)
• Texts: Text descriptions paired with images, stored in corresponding subdirectories
• metadata.parquet: A Parquet file containing structured data for all samples in the split

Data Format

Directory Structure

hf_reevaluate_upload/
├── train/
│   ├── images/
│   │   ├── 000/
│   │   ├── 001/
│   │   └── ...
│   ├── texts/
│   │   ├── 000/
│   │   ├── 001/
│   │   └── ...
│   └── metadata.parquet
├── validation/
│   ├── images/
│   ├── texts/
│   └── metadata.parquet
└── test/
    ├── images/
    ├── texts/
    └── metadata.parquet

Parquet Schema

Each sample in the Parquet files contains the following columns:

Column	Type	Description
image	string	Relative path to the image file
uuid	string	Unique identifier for the artwork
query_text	string	User query-like text
target_text	list[string]	Description text corresponding to the specific image including visual content and metadata information

Text Generation Methods

1. Metadata Portion

The metadata descriptions are constructed by combining multiple metadata fields from the ArtKB knowledge base using different templates. Each template produces a different textual representation of the same metadata information. This results in 5 distinct variants that capture the same facts in different phrasings.

Example fields used:

• Creator/Artist name
• Creation date
• Materials and techniques
• Dimensions
• Current location/Museum
• Object type and classification
• ...

2. Content Portion

The content descriptions are generated automatically using the Salesforce/BLIP2-OPT-2.7B vision-language model. These descriptions capture visual characteristics of the artwork observed directly from the image, such as composition, colors, subjects, and visual elements.

Model: Salesforce/blip2-opt-2.7b

3. Description Texts

The description text descriptions are created by concatenating content portion with metadata protion:

[Content Portion] + [Metadata Portion]

Usage

The dataset can be loaded and used with the Hugging Face datasets library:

from datasets import load_dataset
from IPython.display import display
from PIL import Image
import io

ds = load_dataset("xuemduan/reevaluate-image-text-pairs")

sample = ds["train"][0]

print(sample["uuid"])
print(sample["object_type"])
print(sample["query_text"])
print(sample["target_text"])
display(sample["image"])

Citation

If you use this dataset in your research, please cite this dataset.

Contact

For questions or issues related to this dataset, please email xuemin.duan@kuleuven.be