README.md

metadata

dataset_info:
  features:
    - name: image_name
      dtype: string
    - name: image
      dtype: image
    - name: labelme
      dtype: string
    - name: obb
      dtype: string
  splits:
    - name: train
      num_bytes: 1854583534
      num_examples: 5154
    - name: validation
      num_bytes: 446502719
      num_examples: 1288
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
      - split: validation
        path: data/validation.parquet
license: cc-by-4.0
task_categories:
  - object-detection
language:
  - km
size_categories:
  - 1K<n<10K

Graph Dataset: Image, LabelMe, and OBB

This dataset contains graph/chart images paired with LabelMe JSON polygon annotations and OBB text annotations. Rows are reproducibly divided into 80% training and 20% validation splits.

Dataset Summary

Metric	Count
Matched image/LabelMe/OBB rows	6442
Train rows	5154
Validation rows	1288
Source image files	6442
Source LabelMe JSON files	6442
Source OBB text files	6442
Train row groups	52
Validation row groups	13
Combined parquet size	2194.49 MB

Columns

Column	Type	Description
image_name	string	Source filename without extension
image	image	Image decoded by Hugging Face Datasets
labelme	string	LabelMe annotation serialized as JSON
obb	string	OBB annotation text from the matching .txt file

Load the Dataset

import json

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={
        "train": "data/train.parquet",
        "validation": "data/validation.parquet",
    },
)
row = dataset["train"][0]
image = row["image"]
labelme = json.loads(row["labelme"])
obb_lines = row["obb"].splitlines()

The parquet file is written in bounded row groups by create_hf_parquet_dataset.py to avoid constructing the complete image dataset in memory during export.

Last updated: 2026-05-25