Tibetan STR Dataset for Natural Scene Text

The development of Tibetan scene text recognition systems has been significantly constrained by the absence of large-scale annotated datasets. Unlike well-resourced languages such as English and Chinese, Tibetan STR research lacks standardized evaluation protocols and sufficient training data. To address this critical infrastructure gap, we introduce TibNST, a large-scale dataset specifically designed for Tibetan natural scene text recognition. The TibNST dataset comprises 2,049 meticulously collected Tibetan natural scene images representing diverse real-world scenarios. Data collection encompasses multiple authentic sources including street signs, temple plaques, commercial signboards, and handwritten texts on walls and public displays. Images were captured across various lighting conditions, viewing angles, and environmental settings, resulting in comprehensive coverage of complex backgrounds, illumination changes, rotational distortions, blur effects, and natural interferences commonly encountered in practical applications. Professional annotation was conducted using LabelStudio, with text content transcribed as Unicode encoding sequences by native Tibetan speakers with expertise in traditional and modern scripts. The dataset contains 45,473 characters in total, with text instances ranging from 1 to 92 characters per image, exhibiting diverse character styles with characteristic structural stacking and stroke adhesion features. The annotated samples demonstrate substantial variation in font styles, character sizes, and layout orientations, including stretched text, internal and external ligatures, and diagonal arrangements, making the recognition task considerably more complex than conventional printed text. Detailed dataset analysis is provided in Appendix B.