Robust Reading Competition (RRC) is a prestigious scientific playground, organized by the Center for Computer Ophthalmology (CVC) of the University of autonoma de Barcelona (UAB) Spain, a prestigious research institution in the field of computer vision.
Dr. Dang Minh Tuan - Director of CMC ATI - shared: "We are very happy that the research capacity of the CMC team is affirmed through a globally prestigious playground such as RRC. Proud that in just a short time, the team can achieve a high ranking, competing with big names from developed countries".
Cati-VLM is different from traditional OCR (optical character recognition) in that it does not criticize character extraction, but also understands many layers of information: text content, non- Textual elements (tick, checkbox, chart, signature, formula), layout (page structure, spreadshead, form) and style (font letters, highlights...).
The model can answer intuitive questions raised on document images, similar to ChatGPT, without having to learn each specific form in advance.
On the RRC ranking, CATI-VLM with only 3 billion parameters achieved the highest accuracy in 4/7 sets of data. This achievement proves that CMC's team of engineers are not racing for meaningless parameters but focus on mastering the core of technology, operating effectively on infrastructure suitable for Vietnam.
Mr. Nguyen Trung Chinh - Chairman of the Board of Directors/Chairman of the Executive Committee of CMC Technology Group - emphasized: "This achievement is the sweet fruit of more than 10 years of steadfast investment in R&D of CMC Technology Group, affirming the strategy of mastering Vietnamese technology, associated with the orientation of AI Transformation, Go Global".
Cati-VLM will be applied in the product line of the C. OpenAI ecosystem, including: virtual assistant for CLS reviewing legal documents, CMC SmartDoc - document digital transformation platform, CMC KMS knowledge management system, automatic reporting system for smart offices and new generation Agentic Documents applications.