Map Images With Textual Captions From CommonPool

MapPool contains 75 million potential maps and associated textual captions. It was created by author sraimund and derived from the CommonPool dataset of 12 billion text-image pairs. The dataset was last updated on September 2, 2024.

Use Cases

• Train map detection models using the vision transformer-encoded image features.
• Develop image captioning systems for cartographic content using the textual captions.
• Benchmark classification models against the reported 98.5% validation accuracy for map/non-map separation.
• Fine-tune vision-language models on a domain-specific corpus of map-image-text pairs.

Strengths

• 75 million map-image-text pairs provide significant scale.
• Derived from a source corpus of 12 billion text-image pairs.
• Validation accuracy for map classification reported at 98.5%.

Limitations

• Specific row count, column names, and sample data are unavailable.
• Potential bias from the Internet-sourced CommonPool corpus.
• Image quality, caption relevance, and geographic coverage are unknown.

Provenance

Source

Derived from the CommonPool dataset.

Collection Method

Images encoded by a vision transformer and classified by a support vector machine.

Freshness

Last updated September 2024.