# GloSAT Historical Measurement Table Dataset > GloSAT is a 500-image annotated dataset of historical meteorological measurement logbooks for table structure recognition, with enhanced annotations for headings, headers, table bodies, and coarse segmentation cells. - Authors: Hunter Heidenreich - Published: 2021-09-05 - Updated: 2026-03-26 - Paper: Published (HIP@ICDAR 2021) - Dataset: GloSAT Table Dataset (GitHub) - Dataset: GloSAT Table Dataset (Zenodo) - Source: Bevaya Labs, --- ## TL;DR GloSAT is a manually annotated dataset of 500 historical meteorological logbook page images drawn from nine archival sources spanning the 1700s to the modern day. Beyond standard full-table and cell annotations (ICDAR cTDaR-19 XML and VOC2007 formats), it adds labels for page headings, table headers, table bodies, and coarse segmentation cells that group semantically related data cells. Benchmark experiments using CascadeTabNet show that a simple DBScan-based post-processing step substantially improves cell detection, and that table border type matters more than page style for TSR performance on historical documents. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$: The primary contribution is a new annotated dataset for table structure recognition in historical documents. The paper's central claims rest on the dataset's design, sources, annotation schema, and the open release of splits, code, and model checkpoints. **Secondary:** $\Psi\_{\text{Evaluation}}$: The paper provides benchmark results across five dataset configurations (GloSAT coarse and individual cells, cTDaR-19 tracks B1 and B2, and their aggregation), and introduces a row/col F1 metric that evaluates ruling-line accuracy directly rather than cell-level IoU. ## What is the motivation? Historical data rescue pipelines, such as those digitizing 100,000s of meteorological logbooks in the [GloSAT project](https://www.glosat.org), require table structure recognition applied to scanned images of handwritten and printed measurement records. Existing TSR datasets focus on modern electronic documents (PubLayNet, PubTabNet, TableBank) where annotations can be extracted automatically from PDF metadata, yielding large, richly annotated corpora. Historical document datasets (UNLV, cTDaR-19) are small and limit annotations to full-table bounding boxes and individual cells, omitting contextually important elements: page headings that identify the logbook, table headers that label columns, and coarse groupings of cells indicated by ruling lines. Without annotations for these higher-level structural components, TSR models trained on existing datasets cannot reliably associate data cells with their semantic context, which is precisely what a downstream data rescue pipeline needs. ## What is the novelty? The dataset introduces two categories of annotation beyond the ICDAR standard: 1. **Enhanced contextual annotations** in VOC2007 format: four classes (`heading`, `header`, `table_body`, `background`). A `heading` labels page-level context outside the table region (dates, captions); a `header` labels the column-labeling rows within the table. 2. **Coarse segmentation cells**: annotations that follow the original ruling lines of the logbook, grouping multiple individual data cells into a single coarse cell when the source document groups them by ink lines or whitespace. This preserves the spatial groupings that carry semantic meaning (e.g., groups of weather station IDs from the same region). The evaluation methodology adds a **row/col F1 score** that measures ruling-line accuracy directly: $$ \text{Row/Col F1} = F1\text{ computed on predicted vs. ground-truth ruling lines} $$ where a predicted ruling line matches the ground truth if it falls within $d$ pixels ($d = 20\%$ of the average cell width or height for the horizontal or vertical dimension respectively). This is argued to be more informative than cell IoU for applications such as OCR pipelines that need precise cell boundaries to avoid clipping. The paper also describes the **cTDaR-19 weighted average F1** metric used as the primary cell detection score: $$ \text{Weighted F1} = \frac{\sum\_{i=1}^{4} \text{F1@IoU}\_i \times \text{IoU}\_i}{\sum\_{i=1}^{4} \text{IoU}\_i} $$ where IoU levels are $\\{0.6, 0.7, 0.8, 0.9\\}$. ## What experiments were performed? Two benchmark algorithms are evaluated on five dataset configurations: - **CascadeTabNet (baseline):** Cascade Mask R-CNN with HRNetV2p\_W32 backbone. Due to severe class imbalance (35,555 cells vs. 710 full-table instances in GloSAT), two separate models are trained: one for full-table detection and one for cell detection. The backbone and region proposal network are frozen; only the three cascade heads are fine-tuned using Adam ($\text{lr} = 0.0001$ for TSR, $0.0003$ for detection) until loss stabilizes, typically 60-100 epochs. Learning rate is divided by 3 after epoch 50. - **CascadeTabNet + DBScan post-processing:** After CascadeTabNet cell predictions, 1D DBScan clustering is applied independently to horizontal and vertical ruling lines. Each detected cell contributes two points (start and end coordinate) to each dimension's clustering. DBScan parameters are set heuristically: $\varepsilon$ = half the average cell width/height; $\text{minPts} = 2$. This extrapolates missing cells from the sparse initial detections. For the GloSAT dataset, results from both the coarse and individual cell models are combined as input to the post-processing step. Dataset configurations tested: GloSAT coarse segmentation (129 test / 371 train), GloSAT individual cells (129/371), cTDaR-19 track B1 (150/600), cTDaR-19 track B2 (250/600), and an aggregated split (529/1571). A 75/25 train/test split is used uniformly. Training hardware: NVIDIA Tesla V100 (16 GB). A breakdown of GloSAT performance by table style (bordered, semi-bordered, borderless) and page style (printed, mixed) is also reported for the best-performing model. ## What are the outcomes/conclusions? Post-processing provides substantial gains in all configurations: | Dataset / Configuration | Baseline Weighted F1 | + Post-processing | | :--- | :---: | :---: | | GloSAT coarse cells (manual table det.) | 0.385 | 0.602 | | GloSAT individual cells (manual table det.) | 0.047 | 0.284 | | cTDaR-19 track B1 | 0.084 | 0.161 | | cTDaR-19 track B2 | 0.082 | 0.155 | | Aggregated | 0.071 | 0.181 | Table border type is the dominant performance factor. Bordered tables reach weighted F1 = 0.79; semi-bordered reach 0.31; borderless reach 0.023. Page style (printed vs. mixed) has smaller and partially confounded effects, since mixed-style documents in the GloSAT collection correlate with the bordered table type. The authors conclude that post-processing which encodes structural regularity (rectangular grid assumption) is valuable even when the underlying model has high precision (0.92 at IoU 0.6) but very low recall (0.042). The coarse segmentation annotation variant outperforms individual cell annotation on tables where multiple rows are grouped by ruling lines, suggesting the annotation design should match the downstream use case. The results provide a public baseline for future researchers, but the absolute scores across all configurations remain low, indicating significant headroom for improvement on historical measurement tables. ## Reproducibility ### Models - Architecture: CascadeTabNet (Cascade Mask R-CNN, HRNetV2p\_W32 backbone). Two separate model instances trained: one for full-table detection, one for cell structure recognition. - No new model architecture is introduced; the contribution is dataset and post-processing. - Model checkpoints are released on GitHub alongside the dataset. ### Algorithms - Optimizer: Adam; learning rate $1 \times 10^{-4}$ (TSR) and $3 \times 10^{-4}$ (detection). - Training duration: 60-100 epochs until loss stabilization; learning rate divided by 3 after epoch 50. - Backbone and RPN frozen; only cascade heads trained. - Post-processing: 1D DBScan per dimension; $\varepsilon =$ half mean cell width/height; $\text{minPts} = 2$. ### Data - 500 page images from nine historical archival sources; train/test split 75/25 (371/129). - Sources include UK Met Office, NOAA, University of Reading, Ben Nevis observatory records, and others spanning 1830s to 1970s. - Annotation tools: TTruth (initial grid layout) + Transkribus (visualization and adjustment). - Annotations exported in VOC2007 and ICDAR cTDaR-19 XML formats; two versions (individual cells and coarse segmentation cells). - Dataset, train/test splits, benchmark code, and model checkpoints are all released at the GitHub repository under a BSD open source license. - Dataset is also archived on Zenodo (record 5363457). ### Evaluation - Primary metric: weighted average F1 at IoU $\in \\{0.6, 0.7, 0.8, 0.9\\}$ (cTDaR-19 convention). - Secondary metric: row/col F1 evaluated on ruling lines within $d = 20\%$ of average cell dimension. - The ICDAR cTDaR-19 competition's special rules for cell adjacency and blank cells are not used here; the same metric is applied to both tables and cells for direct comparison. - Comparisons cover automated and manual table region detection variants. - No statistical significance testing or multiple-run averaging is reported. ### Hardware - Training: NVIDIA Tesla V100, 16 GB memory. The IRIDIS High Performance Computing Facility at the University of Southampton. - Inference hardware and latency not reported. ## BibTeX ```bibtex @inproceedings{ziomek2021glosat, author = {Ziomek, Juliusz and Middleton, Stuart E.}, title = {{GloSAT} Historical Measurement Table Dataset: Enhanced Table Structure Recognition Annotation for Downstream Historical Data Rescue}, booktitle = {Proceedings of the 4th International Workshop on Historical Document Imaging and Processing}, series = {HIP '21}, year = {2021}, publisher = {ACM}, doi = {10.1145/3476887.3476892}, } ```