IBEM: A Dataset for Indexing and Search of Mathematical Expressions in Scientific Document Images
Description
IBEM (Indexing and Search of Mathematical Expressions) is a dataset designed for research on mathematical expression indexing, search, and recognition in scientific document images. It contains more than 8000 page images extracted from 600 scientific documents together with ground-truth annotations for both displayed (29596) and inline (137095) mathematical expressions.
Each annotated expression includes its bounding box coordinates, the original LaTeX transcription, several normalized LaTeX representations, and a query-oriented canonical representation suitable for indexing and search. Document-specific LaTeX preambles are also provided to facilitate faithful rendering and interpretation of the annotated expressions.
Besides the page images and annotation files, the dataset includes predefined evaluation partitions as provided for the ICDAR 2021 Competition on Mathematical Formula Detection, and two utility scripts. The script (extract_GT.py) extracts image/LaTeX ground-truth pairs suitable for mathematical expression recognition benchmarks. The second script (extract_subexp.py) performs structure-aware segmentation of mathematical expressions, generating hierarchical sub-expressions from delimiter scopes and symbolic operators. These segments can be used for probabilistic indexing, partial-expression spotting, and other search-oriented applications.
Notes (English)
Each annotated mathematical expression includes the following information:
- Bounding box coordinates (
x_min,y_min,x_max,y_max) (mathematical expression detection, localization, and region extraction). latex: original LaTeX transcription extracted from the source document (ground-truth transcription and rendering).latex_norm: normalized LaTeX representation with standardized formatting and tokenization (mathematical expression recognition, normalization, and benchmarking).latex_expand: macro-expanded LaTeX representation in which document-specific commands are resolved (ground-truth validation and generation of mathematical expression images for recognition tasks).latex_query: query-oriented canonical representation augmented with hierarchical segmentation markers (mathematical expression recognition, probabilistic indexing, sub-expression indexing, and mathematical expression search).- Document preamble: original LaTeX preamble containing document-specific macro definitions (faithful rendering and macro expansion).
In addition, the dataset includes:
- Predefined dataset partitions (training, validation, testing, and reproducible evaluation).
extract_GT.py(generation of image/LaTeX pairs for mathematical expression recognition datasets).extract_subexpressions.py(automatic generation of hierarchical sub-expressions using the delimiters{@,}@,{,}together with the operators@rel,@bin, and@punct, enabling indexing and search of complete expressions and structurally meaningful sub-expressions).
This work has been partially supported by the Ministerio de Ciencia y Tecnología of Spain under the grant TIN2017-91452-EXP (IBEM), by Generalitat Valenciana under predoctoral grant CIACIF/2021/313, by grant PID2020-116813RB-I00 funded by MCIN/AEI/ 10.13039/501100011033, and by grant PID2024-161104OB-C21 funded by MICIU/AEI/10.13039/501100011033 and by Fondo Europeo de Desarrollo Regional (FEDER), European Union.