Wild-OmniDocBench
A 1,350-image benchmark for evaluating document-parsing robustness against real-world captured pages, derived from OmniDocBench with physical printing, deformation and photography.
Wild-OmniDocBench is a benchmark for evaluating document-parsing robustness under real-world captured conditions, rather than clean scanned or digital-born pages. It is derived from OmniDocBench by converting scanned and digital documents into naturally captured images through controlled physical simulation: printing, deformation and photography under varied lighting.
Its 1,350 images introduce realistic artifacts such as geometric distortions (perspective shifts, bends, wrinkles), illumination variation, screen-capture artifacts such as moire patterns and reflections, and environmental interference such as background overlays and shadows, across document types including books, textbooks, papers, slide decks, newspapers and financial reports. Annotations and the evaluation protocol are inherited from OmniDocBench, so a model is evaluated by substituting these images for the original OmniDocBench set and running the official evaluation scripts.
The current release corresponds to OmniDocBench v1.5.