Natural Scene Text Images Across 12 Languages

105,941 images of text in natural scenes. The images cover 12 languages, including 6 Asian and 6 European languages, with line-level quadrilateral bounding box and transcription annotations.

Use Cases

• Train an OCR model to detect text in natural scenes using the provided line-level quadrilateral bounding box annotations.
• Build a multilingual text recognition system leveraging the 12-language coverage, including 6 Asian and 6 European languages.
• Analyze text transcription accuracy across different photographic angles and natural scene contexts present in the images.

Strengths

• Contains 105,941 annotated images, providing a substantial sample for model training.
• Covers a diverse set of 12 languages, split between 6 Asian and 6 European languages.
• Provides line-level quadrilateral bounding box and text transcription annotations for each image.

Limitations

• The dataset is described as a sample of a larger paid dataset, which may limit its completeness.
• Specific image characteristics, geographic origin, and detailed scene descriptions are not provided in the input.
• The input does not specify the distribution of images across the 12 languages, which could indicate class imbalance.

Provenance

Source

Nexdata

Collection Method

Images of text in natural scenes with manual line-level quadrilateral bounding box and transcription annotation.

Freshness

Last updated on 2025-04-24.