unilm
unilm copied to clipboard
LayoutLMv2: DocVQA labeling rules (heuristic matching)
LayoutLMv2 seems to be a great backbone for several document intelligence tasks. Specifically, the performance on docVQA was amazing. However, the dataset was preprocessed by mapping answers to OCR text boxes but it is a closed book that cannot be used. I wonder that it is possible to share the preprocessed dataset? As a leading group for document intelligence, that will contribute to the field.
Any updates on this?