STIPLAR: Scene Text Image Pairs for Korean, Arabic, and Japanese

STIPLAR is a real-world scene text image dataset containing Korean, Arabic, and Japanese text image pairs. The data was collected from MLT-2019 and web sources and is designed for fine-tuning the STELLAR model on low-resource languages. It was created by yongchoooon and last updated on Hugging Face in November 2025.

Use Cases

• Fine-tuning scene text recognition models like STELLAR based on the described real-world image-text pairs.
• Training multilingual OCR systems based on the inclusion of Korean, Arabic, and Japanese text.
• Benchmarking model performance on low-resource languages based on the dataset's stated focus.
• Researching domain adaptation for scene text from web and competition sources as mentioned in the description.

Strengths

• Focuses on three specific low-resource languages: Korean, Arabic, and Japanese.
• Contains real-world scene text images collected from the MLT-2019 competition and web sources.
• Explicitly designed for a specific downstream task: fine-tuning the STELLAR model.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Dataset download location is noted as pending an update, indicating potential access issues.

Provenance

Source

MLT-2019 competition and web sources.

Collection Method

Collected from real-world scenes.

Time Range

null

Freshness

Last updated 2025-11 06 11:32:54; freshness should be verified.

Geography

null