InftyMCCDB-2 dataset
Description
InftyMCCDB-2 dataset is a modified version of InftyCDB-2 which contains mathematical expressions from scanned article pages.
The original dataset has 21,056 math expressions. We remove formulas with matrices and grids, leaving 19,381 formulas. The dataset includes 213 symbol classes, and is split into two sets: training (12551 images), and testing (6830 images) with approximately the same distribution of symbol classes and relation classes. The expressions range in size from a single symbol to more than 75 symbols, with an average of 7.33 symbols per expression.
The original InftyCDB-2 provides ground truth at the symbol level. We extracted connected component bounding boxes, and generated new ground truth for each image using a labeled adjacency matrix (`label graph') representation.
The set of .lg (label graph) ground truth files are provided, along with a .png image for each expression.
Files
IMG.zip
Additional details
References
- @inproceedings{Mahdavi2019LPGAL, title={LPGA : Line-Of-Sight Parsing with Graph-based Attention for Math Formula Recognition}, author={Mahshad Mahdavi and Michael R. Condon and Kenny Davila}, year={2019} }