# TNCR: Table Net Detection and Classification Dataset > A research note on TNCR, a dataset of 9,428 labeled page images spanning 5 structural table classes for table detection and classification in scanned documents. Pages may contain multiple tables. - Authors: Hunter Heidenreich - Published: 2021-06-19 - Updated: 2026-03-16 - Paper: Preprint (arXiv) - Paper: Published (Neurocomputing 2022) - Dataset: TNCR Dataset - Code: TNCR Code & Model Checkpoints - Source: Bevaya Labs, --- ## TL;DR TNCR is a dataset of 9,428 labeled page images collected from publicly available documents (primarily FDA drug labels). Pages may contain multiple tables; the total table count exceeds the page count. Tables are annotated for detection (bounding boxes) and classified into five structural categories: full lined, no lines, merged cells, partial lined, and partial lined merged cells. The authors benchmark several object detection architectures, with Cascade Mask R-CNN using a ResNeXt-101-64x4d backbone achieving the best overall F1 score of 84.4% at IoU 50%:95%. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$: The primary contribution is the TNCR dataset itself, including its five-class taxonomy of table types and the open-source release of images, annotations, and model checkpoints. The paper spends considerable effort describing the data collection pipeline, class definitions, and dataset statistics. **Secondary:** $\Psi\_{\text{Method}}$: The authors train and compare multiple object detection models (Cascade R-CNN, Cascade Mask R-CNN, Cascade RPN, HTC, YOLOv3) with various backbone configurations, establishing baselines for the new dataset. ## What is the motivation? Table detection and classification in scanned document images remains a difficult task due to the diversity of table layouts, image quality variations, and the presence of elements such as merged cells, partial borders, or no borders at all. Existing datasets (ICDAR2013, UNLV, Marmot, TableBank, ICDAR2019) either focus narrowly on table detection without structural classification, are relatively small, or consist of synthetically generated tables from LaTeX or Word. The authors argue that a dataset with real-world images, varying quality, and a structural classification scheme would be a useful resource for the community. ## What is the novelty? The paper introduces a table classification taxonomy that divides tables into five structural categories: 1. **Full lined**: All cell boundaries are drawn with lines, no merged cells. 2. **No lines**: No visible lines delimiting cells. 3. **Merged cells**: Similar to full lined, but containing at least one merged cell. 4. **Partial lined**: Some border lines are missing, but no merged cells. 5. **Partial lined merged cells**: Some border lines are missing, and merged cells are present. The dataset creation process is notable for its semi-automated pipeline. The authors first parsed 875,026 PDF pages from the FDA website (accessdata.fda.gov), then ran a Faster R-CNN model (trained on an initial unbalanced set) to identify 225,154 pages containing tables. They used the detected tables to augment underrepresented classes (no lines, partial lined, partial lined merged cells) and re-partition the dataset. Standard COCO-style evaluation metrics are used: $$ \text{AP} = \frac{\text{TP}}{\text{TP} + \text{FP}} $$ $$ \text{AR} = \frac{\text{TP}}{\text{TP} + \text{FN}} $$ $$ \text{F1} = \frac{2 \times \text{AP} \times \text{AR}}{\text{AP} + \text{AR}} $$ These metrics are computed across IoU thresholds from 50% to 95%. ## What experiments were performed? The authors trained and evaluated five families of object detection models on the TNCR dataset: 1. **Cascade Mask R-CNN** with backbones: ResNet-50, ResNet-101, ResNeXt-101-32x4d, ResNeXt-101-64x4d (various learning rate schedules: 1x, 20e, 2x). 2. **Cascade R-CNN** with the same set of backbones and schedules. 3. **Cascade RPN** with Fast R-CNN and CRPN methods using ResNet-50. 4. **Hybrid Task Cascade (HTC)** with ResNet-50, ResNet-101, ResNeXt-101-32x4d, and ResNeXt-101-64x4d backbones. 5. **YOLOv3** with DarkNet-53 at input scales 320, 416, and 608. All models used SGD with momentum 0.9, weight decay 0.0001, and learning rate 0.02. Images were resized to $1300 \times 1500$ with batch size 16. All models used Feature Pyramid Network (FPN) as the neck. The dataset was split 70%/15%/15% for training, validation, and testing, with the split balanced across classes. Evaluation was performed using precision, recall, and F1 at each individual IoU threshold from 50% to 95% (in 5% increments), as well as the average across IoU 50%:95%. ## What are the outcomes/conclusions? Key results at IoU 50%:95%: | Model | Best Backbone | F1 | |---|---|---| | Cascade Mask R-CNN | ResNeXt-101-64x4d | 0.844 | | Cascade R-CNN | ResNeXt-101-64x4d | 0.841 | | HTC | ResNet-50 (1x) | 0.840 | | Cascade RPN (Fast R-CNN) | ResNet-50 | 0.804 | | YOLOv3 | DarkNet-53 (scale 320) | 0.492 | Cascade Mask R-CNN with ResNeXt-101-64x4d achieved the best performance overall: precision 0.797, recall 0.898, and F1 0.844 at IoU 50%:95%. At the commonly reported IoU 50% threshold, the same model reached an F1 of 0.931. YOLOv3 performed notably worse than the two-stage detectors, particularly at higher IoU thresholds. At IoU 90% and above, YOLOv3's F1 dropped below 0.35 across all input scales, suggesting that the single-stage detector struggled with precise localization. The authors note that ResNeXt-101-64x4d and ResNeXt-101-32x4d backbones suffered from overfitting when used with HTC, while performing well with Cascade R-CNN and Cascade Mask R-CNN. This suggests that the dataset size (approximately 6,500 training images) may be insufficient for the largest model configurations in some frameworks. Limitations worth noting: - The dataset is sourced almost entirely from FDA drug labels, which limits domain diversity. - The class distribution is not balanced; three classes required augmentation from parsed FDA documents. - The paper does not report per-class detection or classification accuracy, making it unclear how well models handle each table type. - No comparison is made against other table detection datasets using the same models. ## Reproducibility ### Models - Architectures: Cascade R-CNN, Cascade Mask R-CNN, Cascade RPN, HTC, YOLOv3 - Backbones: ResNet-50, ResNet-101, ResNeXt-101-32x4d, ResNeXt-101-64x4d, DarkNet-53 - All models use FPN as the feature neck - Pretrained model checkpoints are released on the GitHub repository under MIT license ### Algorithms - Optimizer: SGD with momentum 0.9 and weight decay 0.0001 - Learning rate: 0.02 - Learning rate schedules: 1x (12 epochs) and 20e (20 epochs) for most models - Batch size: 16 - Input resolution: $1300 \times 1500$ - Implementation: MMDetection library for PyTorch (except YOLOv3) ### Data - 9,428 labeled table images across approximately 6,621 document pages - Sources: publicly available documents, primarily from accessdata.fda.gov - Five classes: full lined (2,698), no lines (2,099), merged cells (2,013), partial lined (1,379), partial lined merged cells (1,149) - Split: 70% training, 15% validation, 15% testing (split per class) - Initial collection was unbalanced; three classes were augmented using a Faster R-CNN model to find additional tables from 875,026 parsed FDA PDF pages - Annotations are in COCO format with bounding boxes and class labels - Available for download via SharePoint and Google Drive links on the GitHub repository ### Evaluation - Metrics: precision, recall, and F1 score computed at each IoU threshold from 50% to 95% in 5% increments, plus the mean across 50%:95% - True positive defined as a detection whose bounding box IoU with a ground truth box meets the threshold - No per-class breakdown of detection or classification performance is reported - No cross-dataset evaluation is performed - No error bars, significance tests, or multi-run statistics are reported ### Hardware - Google Colab with 3 Tesla V100-SXM GPUs (16 GB each) and 16 GB RAM - Additionally: a machine with 2x Intel Xeon E5-2680 CPUs and 4x NVIDIA Tesla K20X GPUs - No training time or GPU-hours reported - No inference latency measurements reported ## BibTeX ```bibtex @article{abdallah2022tncr, title={TNCR: Table Net Detection and Classification Dataset}, journal={Neurocomputing}, year={2022}, issn={0925-2312}, doi={10.1016/j.neucom.2021.11.101}, url={https://www.sciencedirect.com/science/article/pii/S0925231221018142}, author={Abdelrahman Abdallah and Alexander Berendeyev and Islam Nuradin and Daniyar Nurseitov} } ```