DocSplit: A Comprehensive Benchmark Dataset and Evaluation Approach for Document Packet Recognition and Splitting

DocSplit is a newly announced benchmark dataset and evaluation framework for the under-explored task of document packet splitting, a critical step in document understanding pipelines aimed at separating mixed multi-page document packets into individual documents. The dataset, introduced in a paper on arXiv (2602.15958v1) by a group of researchers working on visual document understanding, is designed to address the gap in available resources for this specific problem, highlighting its importance in real-world applications such as legal, financial, and administrative workflows where documents are often combined. The dataset contains diverse, heterogeneous document packets, annotated with ground-truth splits and layout information, and is accompanied by an evaluation protocol that benchmarks models on metrics like split accuracy and reconstruction fidelity. The authors argue that while visual document understanding has progressed significantly, the ability to accurately separate and recombine documents remains a bottleneck, especially when documents are scanned or photographed as a single packet, which is common in practice. By providing both data and evaluation methods, DocSplit aims to stimulate focused research and enable comparative studies of new techniques in this domain. Key contributions include: (1) a curated collection of packet images spanning multiple document types and formats; (2) detailed pixel-level annotations for splits and layout boundaries; (3) a set of baseline methods and performance metrics; and (4) open-source code that facilitates reproducible experiments. The dataset is released alongside a benchmark website where researchers can submit results and compare against state-of-the-art baselines, thereby promoting progress in the field. Overall, DocSplit fills a critical niche by offering the first comprehensive resource tailored to document packet recognition and splitting, encouraging the development of robust algorithms that can reliably handle real-world, multi-document inputs.