Reproduce “M$^{6}$Doc: A Large-Scale Multi-Format, Multi-Type, Multi-Layout, Multi-Language, Multi-Annotation Category Dataset for Modern Document Layout Analysis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding and 1 more
The M$^6$Doc dataset is introduced for Document Layout Analysis, featuring 9,080 pages across diverse formats, types, layouts, and languages, annotated with 74 fine-grained categories. A new transformer-based instance segmentation model, TransDLANet, achieves a 64.5% mAP on M$^6$Doc, demonstrating improved generalization to real-world document complexities compared to existing baselines.
Key insights
- Diversity in document formats (scanned/photographed), types, and languages is crucial for training DLA models that generalize effectively to real-world scenarios and varied document structures.
- A rich set of fine-grained annotation labels (74 categories in M$^6$Doc) can be more beneficial for DLA model performance and granular segmentation than merely a larger volume of data with coarser labels.
- Transformer-based architectures with adaptive element matching and iterative refinement can mitigate challenges posed by diverse instance scales and dense/skewed document elements, improving DLA accuracy over anchor-based methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.