# Revisiting Table Detection Datasets for Visually Rich Documents > Open-Tables and ICT-TD: two cleaned and extended table detection datasets addressing noise, annotation inconsistency, and domain coverage gaps in existing TD benchmarks. - Authors: Hunter Heidenreich - Published: 2023-05-08 - Updated: 2026-03-26 - Paper: Preprint (arXiv) - Paper: Published (IJDAR 2025) - Dataset: Open-Tables + ICT-TD Dataset - Source: Bevaya Labs, --- ## TL;DR This paper revisits five existing manually annotated table detection (TD) datasets, cleans their noisy annotations, aligns annotation definitions across them, and merges them into a single larger dataset called Open-Tables. The authors also introduce ICT-TD, a new manually annotated dataset drawn from ICT commodity PDF datasheets, which provides a domain-specific source underrepresented in prior work. Cross-domain baselines show that cleaned data yields 0.6--2.6% higher weighted F1 than training on the noisy originals. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$: The primary contribution is two new or improved datasets (Open-Tables and ICT-TD) released for the table detection community. The paper's dataset construction methodology, annotation rules, and quality analysis are the focal content. **Secondary:** $\Psi\_{\text{Evaluation}}$: The paper builds cross-domain and noise-impact benchmarks, quantifying how annotation quality and domain shift affect model evaluation reliability. ## What is the motivation? Table detection is a foundational step in document understanding pipelines. However, existing public datasets for this task suffer from three problems the authors identify: 1. **Noisy annotations.** Automatically generated datasets like TableBank and PubLayNet contain bounding boxes that are too large, too small, or missing tables entirely. Even manually annotated sets such as ICDAR2013 and TNCR have inconsistent labels due to ambiguous table boundaries (e.g., whether table explanations belong inside the table region). 2. **Inconsistent annotation definitions.** Different datasets draw table boundaries differently, making direct merging unreliable without normalization. 3. **Limited data sources.** Virtually all open TD datasets come from academic publications or government documents. This narrow provenance means that trained models may struggle to generalize to domains with different table layouts, such as industrial product datasheets. ## What is the novelty? The paper's two core contributions are: **Open-Tables.** A merged dataset constructed by cleaning and re-aligning five existing datasets: ICDAR2013, ICDAR2017, ICDAR2019 (modern documents only), Marmot, and TNCR. The authors define two annotation rules applied uniformly across all source sets: - Table boundaries are set by lines; all content bounded by table lines is included, but explanation text outside those lines is excluded. - A table must have at least two rows and two columns. After cleaning (removing bounding boxes that are too large, too small, or mislabeled, and resolving ambiguous samples by the stated rules), the merged result is 8,834 training images, 1,240 test images, and 1,000 validation images. **ICT-TD.** A new dataset sourced from 175,682 PDF files covering 370 different ICT commodity types (network hardware, transceivers, modules, etc.), yielding 3.58 million page images. Random sampling produced 5,000 images containing tables, which were annotated manually under expert guidance. Annotation rules for this domain add a third criterion: a table must describe commodity-specific information, not, for example, a table of contents. The resulting split is 4,000 training / 1,000 test images. ICT domain tables include unique structures not found in academic or governmental datasets: compound tables composed of multiple sub-tables, tables containing figures as cell content, and partially lined tables with domain-specific layouts. The evaluation metric used throughout is the weighted average F1 defined as: $$\text{Weighted Avg. F1} = \frac{\sum\_{i=1}^{4} \text{IoU}\_i \cdot F1@\text{IoU}\_i}{\sum\_{i=1}^{4} \text{IoU}\_i}$$ where IoU thresholds of 80%, 85%, 90%, and 95% are used (stricter than the ICDAR2019 competition setting of 60%--90%). ## What experiments were performed? The authors evaluate four baseline models, selected to span common detection paradigms: - **TableDet**: Cascade R-CNN with transfer learning and table-aware augmentation (two-stage CNN) - **DiffusionDet**: Diffusion-process-based detector with random region proposals - **Deformable-DETR**: Transformer-based detector using deformable attention - **SparseR-CNN**: Learnable sparse proposals with cascade refinement All models are implemented in Detectron2 with a batch size of 16. TableDet uses SGD for 25,000 iterations; the remaining three use AdamW with MultiStepLR for 50,000 iterations. Experiments cover three settings: 1. **In-domain ICT-TD**: train on ICT-TD training set, test on ICT-TD test set. 2. **In-domain Open-Tables**: train on Open-Tables (clean or noisy) training set, test on Open-Tables test set. 3. **Cross-domain**: two directions -- ICT-TD training set evaluated on Open-Tables test set, and Open-Tables training set evaluated on ICT-TD test set. The cross-domain Open-Tables setting is also repeated with the noisy (pre-cleaning) training split to quantify the benefit of noise removal. ## What are the outcomes/conclusions? **In-domain results.** On ICT-TD, Deformable-DETR achieves the best weighted average F1 of 90.3, followed closely by SparseR-CNN (88.9) and DiffusionDet (88.9). On Open-Tables, SparseR-CNN and Deformable-DETR lead at 93.3 and 93.1 respectively. These numbers are high but not directly comparable to benchmarks like ICDAR2013 where near-perfect F1 has been reported, because the authors use stricter IoU thresholds. **Cross-domain results.** Performance drops substantially in both cross-domain directions. Training on ICT-TD and testing on Open-Tables yields weighted F1 of 67.6--75.8 across models. Training on Open-Tables and testing on ICT-TD yields 75.7--80.2. Deformable-DETR shows the best cross-domain generalization in both directions. **Noise cleaning impact.** Training with the cleaned Open-Tables training set consistently outperforms training with the noisy version when tested on the ICT-TD test set, with improvements of 0.6--2.6% weighted F1 across models. The benefit is largest at higher IoU thresholds (90%, 95%), which are most discriminative for bounding box quality. **Limitations the authors acknowledge.** The IoU-based evaluation metrics used throughout are indirect measures of actual information extraction quality. A predicted box that is slightly larger than ground truth but still captures all table content may receive a lower score than a tighter box that misses some rows. The authors suggest that future work should explore metrics more directly tied to downstream information extraction performance. ## Reproducibility ### Models All four baseline models are standard implementations from Detectron2. No custom architecture changes are introduced beyond what is described in the original model papers. Weights for the trained baselines are not released. ### Algorithms | Parameter | TableDet | DiffusionDet | Deformable-DETR | SparseR-CNN | | :--- | :--- | :--- | :--- | :--- | | Optimizer | SGD | AdamW | AdamW | AdamW | | Max iterations | 25,000 | 50,000 | 50,000 | 50,000 | | Base LR | 1e-3 | 1e-5 | 1e-4 | 2.5e-5 | | LR schedule | -- | MultiStepLR | MultiStepLR | MultiStepLR | | Batch size | 16 | 16 | 16 | 16 | ### Data **Open-Tables**: Merges ICDAR2013 (test only, 150 images), ICDAR2017 (1,600 train / 817 test), ICDAR2019 modern (600 train / 240 test), Marmot (2,000 train), and TNCR (4,634 train / 987 test / 1,000 val). After cleaning and merging: 8,834 train / 1,240 test / 1,000 val images. **ICT-TD**: 175,682 ICT commodity PDFs across 370 product types; 3.58M page images rendered at 200 DPI; 5,000 images with tables selected via random sampling and manually annotated. Split: 4,000 train / 1,000 test. Both datasets are publicly available at [HuggingFace](https://huggingface.co/datasets/uobinxiao/open_tables_icttd_for_table_detection) under an Apache-2.0 license. ### Evaluation Weighted average F1 over IoU thresholds 80%, 85%, 90%, 95%. Precision, recall, and F1 at each IoU from 50% to 95% are also reported in appendix tables. No error bars or multi-run statistics are reported. No statistical significance tests are conducted. ### Hardware Not specified in the paper. All experiments are implemented in Detectron2; GPU type, GPU count, and wall-clock training time are not reported. ## BibTeX ```bibtex @article{xiao2025revisiting, author = {Xiao, Bin and Simsek, Murat and Kantarci, Burak and Abu Alkheir, Ala}, title = {Revisiting Table Detection Datasets for Visually Rich Documents}, journal = {International Journal on Document Analysis and Recognition (IJDAR)}, year = {2025}, doi = {10.1007/s10032-025-00527-9}, url = {https://doi.org/10.1007/s10032-025-00527-9} } ```