ROSA: Addressing Text Understanding Challenges in Photographs via ROtated SAmpling
Published in arXiv preprint arXiv:2506.03665, 2025
Visually impaired users rely heavily on assistive VQA systems to read labels, signs, and documents. However, photographs taken in natural settings often feature tilted or upside-down text that degrades VLM performance. ROSA introduces an effective rotated sampling strategy that restores OCR and QA accuracy without retraining.