maianh511/classified_charts

Data Card: Chart Image Dataset (Crawled from arXiv)

1. Overview

This dataset contains chart images automatically collected from research papers on arXiv.org. Charts were extracted from PDF pages and classified by type, resulting in a structured collection organized into 8 chart categories.

2. Data Collection Pipeline

Stage	Method	Output Count
Source papers	Automated download via arXiv API	15,000+ PDFs
Chart extraction	PDF page + bounding box extraction (PyMuPDF)	46,911 chart images
Chart classification	Multi-model classification (Gemini) with confidence thresholding	32,364 labeled images

3. Chart Type Distribution

Chart Type	Count	Share
line	12,930	40.0%
scatter	6,278	19.4%
bar	5,745	17.8%
heatmap	4,073	12.6%
histogram	1,006	3.1%
box	880	2.7%
pie	835	2.6%
area	617	1.9%
Total	32,364	100%

4. Metadata & Schema

Each record includes:

• id: Unique identifier
• image: Chart image extracted from the source PDF
• label: Chart type category (one of the 8 classes above)

5. Notes

• Charts were extracted directly from arXiv paper pages, so image quality and resolution vary depending on the original PDF.
• Classification labels were assigned automatically (Gemini-based) and may contain minor noise; a confidence threshold was applied during labeling.
• Of the 46,911 raw extracted charts, 32,364 were successfully classified into one of the 8 defined categories.