Datasets:

Geologi
/

imregis

language:
  - pt
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-text
tags:
  - image-captioning
  - multimodal
  - oil-and-gas
  - geosciences
  - portuguese
configs:
  - config_name: images
    data_files:
      - split: train
        path: data/images/train-*.parquet
  - config_name: annotations
    data_files:
      - split: train
        path: data/annotations/train-*.parquet

ImREGIS

ImREGIS (Image-REGIS) is a Portuguese multimodal image-text dataset, automatically extracted from the REGIS collection — a set of technical documents, theses, and reports from the Oil & Gas (O&G) and Geosciences domain.

The dataset contains over 439,000 unique images and 581,000 image-text pairs, extracted from more than 20,000 PDF documents, combining captions (descriptive text placed near the image) and descriptions (textual references to the image scattered throughout the document body).

Structure

The dataset is split into two tables (configs) linked by the image_id column.

A single doc_id may have more than one image associated with it (documents with multiple figures/pages).

`images`

One row per unique image (deduplicated by image_id).

column	type	description
`image_id`	string	unique image identifier (join key)
`doc_id`	string	source document identifier
`width`	int32	width in pixels
`height`	int32	height in pixels
`format`	string	original format
`image`	binary (bytes)	raw image content

`annotations`

One or more text rows per image.

column	type	description
`row_id`	string	unique annotation row identifier
`doc_id`	string	source document identifier
`image_id`	string	identifier of the associated image (key for `images`)
`type`	string	`caption` or `description`
`text`	string	text content
`lang`	string	language

How to load

from datasets import load_dataset

images = load_dataset("Geologi/imregis", "images", split="train")
annotations = load_dataset("Geologi/imregis", "annotations", split="train")

Given the dataset size (250GB+), streaming is recommended instead of loading everything into memory:

images = load_dataset("Geologi/imregis", "images", split="train", streaming=True)
annotations = load_dataset("Geologi/imregis", "annotations", split="train", streaming=True)

Joining images and text

For local use (if it fits in memory/disk), via pandas, always join on image_id:

import pandas as pd

df_images = images.to_pandas()
df_annot = annotations.to_pandas()

df = df_annot.merge(df_images, on="image_id", how="left", suffixes=("", "_img"))

For streaming use, perform the join on demand, for example by building an in-memory index of images (if the images table fits) and iterating over annotations:

from datasets import load_dataset

images_index = {
    row["image_id"]: row["image"]
    for row in load_dataset("Geologi/imregis", "images", split="train", streaming=True)
}

for row in load_dataset("Geologi/imregis", "annotations", split="train", streaming=True):
    image_bytes = images_index.get(row["image_id"])
    # use row["text"], row["type"], row["lang"], and image_bytes

Decoding the image bytes

The image column contains the raw bytes of the original file. To open it as an image:

from PIL import Image
import io

img = Image.open(io.BytesIO(images[0]["image"]))
img.show()