# Parsing Table Structures in the Wild (WTW) > A research note on the WTW dataset and Cycle-CenterNet, which address table structure parsing from natural images with deformation, bending, and occlusion. - Authors: Hunter Heidenreich - Published: 2021-09-06 - Updated: 2026-03-16 - Dataset: WTW Dataset - Paper: Preprint (arXiv) - Source: Bevaya Labs, --- ## TL;DR This paper introduces the Wired Table in the Wild (WTW) dataset, a collection of 14,581 images of tables captured in natural scenes (photos, scans, web pages) with full cell coordinate and row/column annotations. Alongside the dataset, the authors propose Cycle-CenterNet, which extends CenterNet with a cycle-pairing module to simultaneously detect cells and group them into structured tables, achieving a 24.6-point absolute improvement in TEDS over baselines on WTW. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$ The primary contribution is the WTW dataset itself, which fills a gap in available benchmarks for table structure parsing (TSP) outside of clean document images. The paper spends considerable effort describing the data collection, annotation protocol, challenging cases, and evaluation protocol. **Secondary:** $\Psi\_{\text{Method}}$ The paper also proposes Cycle-CenterNet with a novel cycle-pairing module and pairing loss, which represents a meaningful methodological contribution for cell detection and grouping. ## What is the motivation? Existing table structure parsing datasets (ICDAR-2013, PubTabNet, TableBank, SciTSR) focus almost exclusively on well-aligned tabular images from digital or scanned PDF documents. These images have clean backgrounds and clear, axis-aligned table structures. In practice, tables are frequently captured by hand-held cameras, introducing challenges like non-rigid deformation, bending, tilting, occlusion, and cluttered backgrounds. The authors observe that existing TSP methods, which rely on assumptions about well-aligned inputs, fail on these real-world images. There was no large-scale dataset to train or evaluate models for this more practical setting. The authors also note that wireless tables in natural images are particularly difficult for human annotators due to the lack of reference lines, so they focus specifically on wired tables. ## What is the novelty? ### WTW Dataset The dataset contains 14,581 images drawn from three sources: natural scene images (50%), archival document images (30%), and printed document images (20%). Each image is annotated with table IDs, cell coordinates (using inner table lines for localization, following ICDAR 2019 conventions), and row/column information. The dataset is split into 10,970 training and 3,611 testing samples. WTW covers seven challenging cases that existing datasets largely miss: 1. Inclined tables 2. Curved tables 3. Occluded or blurred tables 4. Extreme aspect ratio tables 5. Overlaid tables 6. Multi-color tables 7. Irregular tables ### Cycle-CenterNet Built on top of CenterNet with a DLA-34 backbone, the model adds a cycle-pairing module with two branches: **Center-to-Vertex branch:** Predicts offsets from each cell center to its four vertices. For center point $P = \\{x\_C, y\_C\\}$ and vertices $V = \\{x\_V, y\_V\\}$: $$ \begin{cases} \Delta x\_{C\_{ik}} = x\_{C\_i} - x\_{V\_{ik}} \\\\ \Delta y\_{C\_{ik}} = y\_{C\_i} - y\_{V\_{ik}} \end{cases}, \quad i=1:N^C, \; k=1:4 $$ **Vertex-to-Center branch:** Predicts offsets from each common vertex to the centers of surrounding cells: $$ \begin{cases} \Delta x\_{V\_{ik}} = x\_{V\_i} - x\_{C\_{ik}} \\\\ \Delta y\_{V\_{ik}} = y\_{V\_i} - y\_{C\_{ik}} \end{cases}, \quad i=1:N^V, \; k=1:4 $$ The mutual-directed (cycle) relationship between centers and vertices enables grouping cells through shared vertices at cell intersections. **Pairing loss:** Rather than applying loss functions directly on the output maps, the authors define a pairwise loss over center-vertex pairs: $$ L\_p = \sum\_{c,v} \omega(P\_{cv}) \left( \lambda\_{cv} L\_{cv} + \lambda\_{vc} L\_{vc} \right) $$ where $L\_{cv}$ and $L\_{vc}$ are $l\_1$ losses on the predicted offsets, and $\omega(P\_{cv})$ is a dynamic weighting function: $$ \omega(P\_{cv}) = 1 - \exp(-\pi D\_{cv}) $$ Here $D\_{cv}$ measures the regression error for each center-vertex pair, so that pairs with larger errors receive higher weight during training. The overall training loss combines the keypoint loss, offset loss, and pairing loss: $$ L\_{det} = L\_k + \lambda\_{off} L\_{off} + L\_p $$ A post-processing parsing module then recovers row/column information by splitting each cell into bounding edges, merging them into horizontal and vertical lines, and sorting/indexing them. ## What experiments were performed? ### Evaluation on WTW The authors evaluate physical structure accuracy (cell coordinate precision, recall, F1 at IoU $= 0.9$) and logical structure accuracy (adjacency relation precision, recall, F1 at IoU $= 0.6$, plus TEDS). Baselines include four object detectors (Faster-RCNN, TridentNet, Cascade-RCNN, CenterNet) with heuristic grouping, as well as two existing TSP methods (Split+Heuristic, CascadeTabNet) retrained on WTW. Key results on WTW: | Model | Phys. F1 | Adj. F1 | TEDS | |:---|:---:|:---:|:---:| | Split+Heuristic | 3.4 | 27.6 | 26.0 | | CenterNet (baseline) | 73.1 | 84.8 | 58.7 | | Cycle-CenterNet | 78.3 | 92.4 | 83.3 | The existing document-oriented TSP methods perform poorly on WTW, confirming that this dataset introduces genuinely new challenges. Cycle-CenterNet improves over the CenterNet baseline by 5.2 points on physical F1, 7.6 points on adjacency F1, and 24.6 points on TEDS. ### Cross-dataset evaluation On **ICDAR 2013**, Cycle-CenterNet achieves 91.7% adjacency F1, only 1.3% behind the top-ranked Split+Heuristic, and reaches 98.0% on the wired-table subset. On **ICDAR 2019 Track B2**, Cycle-CenterNet improves the weighted-average F1 by 15.2 points over the previous best reported result. These results suggest that training on WTW generalizes well to simpler document-oriented benchmarks. ## What are the outcomes/conclusions? The WTW dataset establishes a new benchmark for table structure parsing in natural, unconstrained settings. The authors demonstrate that methods designed for clean document images (Split+Heuristic, CascadeTabNet) largely fail on wild images, confirming the need for both new data and new approaches. Cycle-CenterNet's cycle-pairing mechanism provides an effective way to jointly detect cells and learn their grouping structure, avoiding the need for heuristic post-processing rules typical of prior work. The dynamic pairing loss focuses training on the hardest center-vertex pairs. Limitations worth noting: the dataset focuses exclusively on wired tables, leaving wireless tables in natural images as an open problem. The model weights are not publicly released (the authors note the model is used in Alibaba's online business software). The paper does not report results on merged or spanning cells in detail, nor does it discuss computational cost or inference latency. ## Reproducibility ### Models - Architecture: CenterNet with DLA-34 backbone, extended with the cycle-pairing module producing two additional 8-channel output maps ($CV\_{map}$ and $VC\_{map}$). - Pretrained on COCO, then trained on WTW. - Model weights are **not publicly released**. The authors note the model is deployed in Alibaba's online business software. An online testing demo was mentioned but may not be currently available. ### Algorithms - Optimizer and schedule: initial learning rate $1.25 \times 10^{-3}$, decayed to $1.25 \times 10^{-4}$ at epoch 90 and $1.25 \times 10^{-5}$ at epoch 120. Total: 150 epochs. - Batch size: 32 per GPU (8 GPUs, so effective batch size of 256). - Input resizing: max side resized to 1024, short side scaled proportionally. - Loss hyperparameters: $\lambda\_{cv} = 1.0$, $\lambda\_{vc} = 0.5$. - Post-processing: edge splitting, line merging, and row/column indexing (pseudo-code in supplementary materials). ### Data - **WTW dataset:** 14,581 images total (10,970 train / 3,611 test). Sources: natural scenes (50%), archival documents (30%), printed documents (20%). - Annotations: cell coordinates (polygonal, using inner table lines), row/column indices, table instance IDs. - Sensitive information (names, phone numbers) has been erased. - The dataset is available via GitHub with a download link to the Tianchi Aliyun platform. - License: CC-BY-NC-4.0. ### Evaluation - **Physical structure:** Precision, Recall, F1 at IoU threshold of 0.9 (stricter than typical object detection). - **Logical structure:** Cell adjacency Precision, Recall, F1 at IoU threshold of 0.6; TEDS (Tree-Edit-Distance-based Similarity). - Cross-dataset evaluation on ICDAR 2013 and ICDAR 2019 Track B2 using their respective official metrics. - Baselines (Split+Heuristic, CascadeTabNet) retrained on WTW with author-provided hyperparameters. - No error bars, significance tests, or multi-run statistics reported. ### Hardware - Training: 8 NVIDIA GTX 1080Ti GPUs. - Total GPU-hours, inference latency, and memory requirements are not reported. ## BibTeX ```bibtex @inproceedings{long2021parsing, title={Parsing Table Structures in the Wild}, author={Long, Rujiao and Wang, Wen and Xue, Nan and Gao, Feiyu and Yang, Zhibo and Wang, Yongpan and Xia, Gui-Song}, booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, year={2021} } ```