AI BriefingUpstage Sweeps 4 Categories at ICDAR Robust Reading Competition
Key point
Upstage swept 4 categories at the ICDAR Robust Reading Competition, the world's most prestigious AI OCR competition, proving its global technological capability.
Details
Upstage took first place in 4 categories—HierText-1/2, VQAonBD, and IHTR—at the ICDAR Robust Reading Competition hosted by the International Association for Pattern Recognition (IAPR), proving its world-class OCR technology. In this competition, Upstage recorded the highest scores in competition with global big tech companies such as Amazon, NVIDIA, Alibaba, and Huawei.
Key achievements are as follows:
- IHTR: Built the best-performing model for 10 Indian regional languages with which it had no prior experience
- HierText: Swept both 1st and 2nd place in the hierarchical text detection and recognition category
- VQAonBD: Achieved overwhelming results in the visual question answering field, which involves extracting data from document images and performing complex calculations
The technical differentiator lies in improving upon existing detection methods. Previously, word regions were shrunk significantly during training to avoid overlap between regions, but the Upstage team introduced a method of inserting gaps between word regions, greatly improving the similarity (tightness) between predicted boxes and ground-truth boxes.
Based on its proven technological capabilities, Upstage supports companies' digital transformation and AI innovation through AI Pack, its API series, and AskUp Biz (AskUp Doc, Web, Slack) for business.
