Tables have been an ever-existing structure to store data. There exist now diferent approaches to store tabular data physically. PDFs, images, spreadsheets, and CSVs are leading examples. Being able to parse table structures and extract content bounded by these structures is of high importance in many applications. In this paper, we devise TableParser, a system capable of parsing tables in both native PDFs and scanned images with high precision. We have conducted extensive experiments to show the eficacy of domain adaptation in developing such a tool. Moreover, we create TableAnnotator and ExcelAnnotator, which constitute a spreadsheet-based weak supervision mechanism and a pipeline to enable table parsing. We share these resources with the research community to facilitate further research in this interesting direction.
of the labeling complexity through this weak supervision
alleviated through a novel weak supervision approach
that automatically generates training data from
structures, DocParser [
0000-0003-2379-1506 (S. X. Rao); 0000-0002-9409-4401
the literature about table processing in PDFs, namely,
table detection, table structure parsing/recognition [
1Table detection is a task to draw the bounding boxes of tables ditional) identification of the structural (row and column layout) information of tables. We distinguish between bottom-up and topdown approaches in table structure detection. Bottom-up typically refers to structure detection by recognizing formatting cues such as text, lines, and spacing, while top-down entails table cell detection
Table detection is a popular task with a large body of
literature, table structure parsing and table recognition
were revisited2 after the pioneering work of [
The DL-based methods in [
However, a key problem in training DL-based systems via a Python library on Excel (PyWin324) to leverage
is the labeling complexity of generating high-quality in- the structured information stored in the spreadsheets.
domain annotations. More generally, an essential limit- TableParser is trained with 16’041 Excel-rendered tables
ing factor is the lack of large amounts of training data. using detectron2 ([21, 22]) and fine-tuned with 17
highEforts have been put into generating datasets to enable quality manual annotations in each domain. We have
tasks with weak supervision. TableBank [20] is built conducted extensive experiments of domain adaptation.
upon a data set of Word and LaTeX files and extracts Finally, we evaluate diferent TableParsers in two
doannotations directly from the sources. They use 4-gram mains and make the following observations:
BLEU score to evaluate the cell content alignments.
However, the table layout structure is not of particular focus 1. In general, domain adaptation works well with
in TableBank. PubTabNet [
In this paper, we devise TableParser with inspiration 73.28% on table rows and columns, respectively.
from DocParser, due to its flexibility in processing both ta- 3. We have tested our HistoricalTableParser on
bles and more general documents. We demonstrate that scanned tables in both historical
(mediumTableParser is an efective tool for recognizing table quality, scan-based) and modern tables.
Overstructures and content. The application of TableParser to all, HistoricalTableParser works better than
Moda new target domain requires newly generated training ernTableParser on tables stored in image scans.
data. Depending on the target domain, we specify two 4. Interestingly, we find that ModernTableParser
TableParsers: ModernTableParser fine-tuned with na- built on top of DocParser [
TableAnnotator (Figure 1) which eficiently assists developers in visualizing the output, as well as help users to generate high-quality human annotations.3 To gener
We are willing to open source the ZHYearbook-Excel
dataset, TableAnnotator, TableParser system, and its
pipeline to the research communities.5 Moreover, we
2Some recent works on Cascade R-CNN [
3For a live demo of table annotations using our annotation tool, refer to the video under https://github.com/DS3Lab/TableParser/ blob/main/demo/2021-06-15%2002-05-58.gif.
4https://pypi.org/project/pywin32/ (last accessed: Sep. 30, 2021).
5The source code, data, and/or other artifacts for the complete TableParser pipeline have been made available at https://github.com/
DS3Lab/TableParser.
Weak Supervision
Document Rendering
Excel Annotator
Annotations
Parsed Table Structure DocParser
MWS
HistoricalTableParser
2.2.1. TableAnnotator. 1. We present TableParser which is a robust tool for parsing modern and historical tables stored in native PDFs or image scans. 2. We conduct experiments to show the eficacy of domain adaptation in TableParser. 3. We contribute a new pipeline (using ExcelAnnotator as the main component) to automatically generate weakly labeled data for DL-based table parsing. 4. We contribute TableAnnotator as a graphical interface to assist table structure understanding and manual labeling. 5. We open-source the spreadsheet weak supervision dataset and the pipeline of TableParser to encourage further research in this direction.
composed of two parts: image panel (left) and document tree (right). In the code repository7, there is a manual describing its functionalities in details. In a nutshell, annotators can draw bounding boxes on the left panel and create their entities and relationships on the right.
In Figure 1, the highlighted bounding box (the red thick
contour on the left) corresponds to the table_cell on the
2. TableParser System second row and second column, indexed by 1-1, 1-1 (the
blue highlight on the right). Note that TableAnnotator
2.1. Problem Description is versatile and can be used to annotate not only tables,
but also generic documents. The output of the tree is in
Following the hierarchical document parsing in Doc- JSON format, as shown in the following code snippet.
Parser, our objective is to generate a hierarchical struc- 1 [{"id": 28,
ture for a table which consists of the entities (table, tabu- 2 "category": "table_cell",
lar, table_caption, table_row, table_column, table_footnote) 34 ""rporwo_prearntgiee"s"::[1"1,-11],,1-1",
and their relations in the document tree. 5 "col_range": [
Our ultimate goal of table structure parsing is (1) to es- 6 "parent": 9},
tablish row/column relationships between the table cells, 7 {"id": 29,
and (2) post-process the established structure and cell 8 "category": "box",
content (e.g., with PDFMiner6 or OCR engines) to enable 109 ""bpbaogxe"":: 0[
6https://pypi.org/project/pdfminer/ (last accessed: Nov. 11, 2021).
annotation-SDU-AAAI22. (a) Example worksheet from ZHYearbook-Excel-WS. (b) Annotations with DeExcelerator. (c) Representing bounding boxes in Excel.
(d) Visualization of bounding boxes with TableAnnotator. 2.2.2. ModernTableParser. We utilize DeExcelerator to categorize the content, such that we can diferentiate among table captions, table footWe train ModernTableParser using the data generated notes and tabular data and create a correct auxiliary file by weak supervision signals from Excel sheets and fine- to each PDF containing the structural information of the tuned by high-quality manual annotations in this domain. represented table(s). Illustrated in Figure 3 (b), in this In Figure 2, we show the system design following the case we annotate the table caption and footnote as ‘meta’, underlying components of DocParser.8 We denote the and mark the range of content with ‘content’ and ‘empty’. model that produces ModernTableParser as M1. We use PyWin32 in Python to interact with Excel, so that intermediate representations like Figure 3 (c) can be creWeak Supervision with ExcelAnnotator. Now we ated to retrieve entity locations in the PDF rendering. present the crucial steps in generating weak supervision Concretely, we mark neighboring cells with distinct col(the model in Figure 2) for TableParser. These steps ors, remove all borders, and set the font color to white. are mainly conducted by ExcelAnnotator in Figure 2 (left). To summarize, ExcelAnnotator detects spreadsheet metaTake a worksheet-like Figure 3 (a) from our ZHYearbook- data and cell types, as well as retrieves entity locations Excel-WS dataset (cf. Section 3), where we see caption, via intermediate representations. Finally, we are able to tabular, and footnote areas. We subsequently use DeEx- load the annotations into TableAnnotator to inspect the celerator [23] to extract relations from the spreadsheets. quality of weak supervision (Figure 3 (d)).
the DocParser paper [
The model structure (Mask R-CNN) can also be found here.
(a) Bad quality of OCR (left). (b) Good quality of OCR (left). 2.2.3. HistoricalTableParser. bounding boxes of individual cells. In Figure 2 (lower right), we show the system design We use the OCR engine from Google Vision API to rec- by adding an OCR component and a fine-tuning compoognize the text bounding boxes. Then we convert bound- nent for domain adaptation. We denote the model that ing boxes into the input format TableParser requires. produces HistoricalTableParser as M2. Take Figure 5 (a) Now we are able to manually adjust the bounding boxes as input, TableParser can produce a parsed layout-like in TableAnnotator to produce high-quality annotations. Figure 5 (b) which can be combined with the OCR boundNote that the quality of OCR highly depends on the ta- ing boxes in the subsequent steps and export as a CSV ble layout (see (a) vs. (b) in Figure 4), we often need to adjust the locations of bounding boxes and redraw the ifle (Figure 5 (c)).9
For domain adaptation, we assume that an outof-domain model performs worse than an in-domain model in one domain. Namely, we would expect ModernTableParser to work better on Excel-rendered PDFs or tables created similarly; on the contrary, we would expect HistoricalTableParser to perform better on older table scans.
an updated backend implementation. Namely, we utilize
Detectron2 to apply an updated version of Mask R-CNN
[25]. For technical details of Mask R-CNN, we refer to
DocParser [
4.1.1. Training Procedure: Weak Supervision + Fine-Tuning.
ZHYearbooks-Excel. We create three datasets All neural models are initialized with weights trained on from this source: ZHYearbooks-Excel-WS, ZHYearbooks- the MS COCO dataset. We first pretrain on the weak suExcel-FT, and ZHYearbooks-Excel-Test, with 16’041, 17, pervision data ZHYearbook-Excel-WS for 540k iterations, and 20 tables in each set. On average, it takes 3 minutes then fine-tune on our target datasets ZHYearbook-Excel30 seconds for an annotator to produce high-quality anno- FT and EUYearbook-OCR-FT for M1 and M2, respectively. tations of a table. The manual annotations are done with We then fine-tune each model across three phrases for automatically generated bounding boxes and document a total of 30k iterations. This is split into 22k, 4k, 4k tree as aid. iterations, respectively. The performance is measured
ZHYearbooks-OCR. We create the dataset every 500 iterations via the IoU with a threshold of 0.5. ZHYearbook-OCR-Test, with 20 tables. On average, it We train all models in a multi-GPU setting, using 8 GPUs takes 2 minutes and 45 seconds to annotate a table with with a vRAM of 12 GB. Each GPU was fed with one imthe similar annotation aids mentioned above. age per training iteration. Accordingly, the batch size
EUYearbooks-OCR. We create two datasets from per training iteration is set to 8. Furthermore, we use this source: EUYearbook-OCR-FT and EUYearbook-OCR- stochastic gradient descent with a learning rate of 0.005 Test, with 17 and 10 tables, respectively. Note that these and learning momentum of 0.9. datasets contain various languages like Hungarian and German, with various formats depending on the language. 4.1.2. Parameter Settings.
On average, it takes 8 minutes and 15 seconds to annotate During training, we sampled randomly 100 entities from a table with the similar annotation aids mentioned above. the ground truth per document image (i.e., up to 100
Miscellaneous historical yearbooks. We ran Mod- entities, as some document images might have less). In ernTableParser and HistoricalTableParser on Chinese (in Mask R-CNN, the maximum number of entity predictions Simplified Chinese) and South Korean historical year- per image is set to 100. During prediction, we only keep books (in Classical Chinese) and inspect their outputs entities with a confidence score of 0.5 or higher. qualitatively (see Section 5.2).
Human labeling eforts. We observe a large variance in labeling intensity across the datasets. The 5. Results and Discussion EUYearbooks-OCR datasets require more corrections per table compared to the datasets of modern tables. More- Here, we evaluate the performance of TableParser in two over, they also require more iterations of human annota- domains quantitatively and qualitatively. tions with heuristics as aid.
ular data in Figure 5 (d) using the best model from its model zoo (PubLayNet/faster_rcnn_R_50_FPN_3x). Input and annotated figures of original size can be found under https://github.com/DS3Lab/ TableParser/tree/main/figures.
the object detection/classification tasks. The metric we report is Average Precision (AP), which corresponds to an Intersection over Union rate of IoU=.50:.05:.95.10 IoU ranges from 0 to 1 and specifies the amount of overlap between the predicted and ground truth bounding box.
It is a common metric used when calculating AP.
10We refer readers to https://cocodataset.org/#detection-eval for
more details on the evaluation metrics (last accessed: Nov. 1, 2021).
Performances in various domains. As we discussed run M1 (fine-tuned by modern tables) on
EUYearbookin Section 2, we have developed ModernTableParser to OCR-FT (column (7) in Table 1), its performance is worse
parse tables with input images rendered by Excel (M1). than fine-tuning; and if we run M2 (fine-tuned by
hisThen, to work with historical tables in scans, we adapt torical tables) on ZHYearbook-Excel-FT (column (4) in
the pretrained TableParser by fine-tuning it on scanned Table 1), it performs worse than fine-tuning.
Interestdocuments (M2). Now, we present the performances of ingly, if we compare the performance of M2 on modern
M1 and M2 in two diferent domains in the following tables (column (4) in Table 1) with the performance of
aspects: M1 on historical tables (column (7) in Table 1), we clearly
1. (P1) the performances on fine-tuning sets on M1 see that the latter has a better performance in all other
and M2 in Table 1; categories than the class of table_row. This can be
explained by the fact that the model trained on modern
2. (P2) the performances on fine-tuning sets as test tables is robust in annotating historical tables, at least on
sets on M1 and M2 in Table 1;11 the column level. We see this in Figures 9 and 10, where
3. (P3) the performances on three test sets from two ModernTableParser clearly performs better. However,
domains on M1 and M2 in Table 2. the algorithm has problems in delineating narrow
(P1) & (P2). We want to study the impact of fine- and less clearly separated rows. This could be due to
tuning of a pretrained model (using a large body of tables the setting of the maximum number of entities being 100
generated by weak supervision signals). The instances when predicting per table (Section 4.1).
used to fine-tune must be high-quality in-domain data. (P3). In Table 2, we show the performances of
Concretely, we create in-domain annotations for mod- three test sets from two domains (Excel-rendered PDFs
ern tables (rendered by Excel) and historical tables (from and historical scans), namely, ZHYearbook-Excel-Test,
scans) with high human eforts assisted by automatic ZHYearbook-OCR-Test, and EUYearbook-OCR-Test. We
preprocessing: ZHYearbook-Excel-FT and EUYearbook- see that M2 which is fine-tuned by historical scans
perOCR-FT, each with 17 tables. Note that the latter has forms worse than M1 on both ZHYearbook-Excel-Test
much denser rows and columns than the former (see and ZHYearbook-OCR-Test. Vice versa, M1 that is
finethe tables in Figures 3 (a) vs. 5 (a) for an illustration). tuned by Excel-rendered PDFs performs worse than M2
It is apparent from Table 1 that the AP performance on EUYearbook-OCR-Test. This suggests that domain
of models on the fine-tuning sets is highly optimized adaptation by fine-tuning the pretrained TableParser
(columns (3) and (8) in Table 1), and it should be better with in-domain high-quality data works well.
than using those datasets as test sets. This means, if we Additionally, if we compare the Δ |( 1− 2)| under
each test set (e.g., the diferences of columns (3) and (4),
of (7) and (8), of (11) and (12) in Table 2), the Δ on
11This means we evaluate the performance of M1 on the
finetuned set for M2 (as a test set for M1) and vice versa.
*-OCR-Test in all categories is smaller than ZHYearbook- The success of DL has marked the revisiting of
taExcel-Test, with M1 already achieving medium-high per- ble structure parsing by [
In Figures 7, 8, 9, and 10 in Appendix C, we show the qualitative outputs of ModernTableParser and HistoricalTableParser on various types of inputs.12 The quality of structure parsing varies across inputs, but overall, the quality is high. Even if we simply use ModernTableParser to parse old scans, it achieves a moderate performance, sometimes better than HistoricalTableParser (see Figures 9 and 10). This substantiates our claim that knowing the table structure (caption, tabular, row, column, multi-cell, etc.) is of foremost importance for parsing tables. We see that the performance of LayoutParser is quite poor on the tabular data in Figure 5 (d) using the best model from its model zoo (PubLayNet/faster_rcnn_R_50_FPN_3x).
only supports single-thread processing and only runs un6. Related Work der Windows. But with 20 VMs, we managed to process a large amount of files. This is a one-time development Table Annotation. TableLab [26] provides an active cost. On average – on the fastest machine used (with learning based annotation GUI for users to jointly opti- 16 GB memory, 6 cores, each of 4.8GHz max (2.9 base)) – mize the model performance under the hood. Layout- it took 15.25 seconds to process one document (a workParser [24] has also promoted an interactive document sheet in this case). To fine-tune a pretrained TableParser annotation tool13, but the tool is not optimized for table with 17 images, it takes 3-4 hours to fine-tune the model annotations. with 30k iterations.
12Input and annotated figures of original size can be found under https://github.com/DS3Lab/TableParser/tree/main/figures.
13See https://github.com/Layout-Parser/annotation-service (last format. Because tables sometimes come with row/colaccessed: Nov. 1, 2021). umn sums in the rendered format, this functionality can
Based on our findings, we will further improve the parsing performance on table row/column/cell. Besides, we plan to enable a CSV-export functionality in TableParser, which allows users to export a CSV file that attends to both bounding boxes generated by the OCR’ed and the hierarchical table structure. We will also benchmark this functionality against human eforts. Another practical functionality we add to facilitate users’ assessment of table parsing quality, is that we enable TableParser to compute row and column sums when exporting to the CSV help users to assess their manual eforts in post-editing the CSV output. We also plan to conduct perturbation tests of table structures and quantify the robustness of our models in those scenarios. These exercises will be highly valuable because, as we see in Figure 7, we often encounter scan images of tables where the rectangle structures cannot be maintained (the upper right corner).
This brings us to another interesting research direction:
how to eficiently annotate the non-rectangle elements in
a table, e.g., [30] have provided the benchmarking dataset
and method for parsing tables in the wild. Finally, we
would like to benchmark TableParser using the
popular benchmarking datasets such as ICDAR-2013,
ICDAR2019, TableBank, and PubTabNet. Note that since we
develop TableParser on top of the DocParser [
table structures and its components: TableAnnotator, TableParser (Modern and Historical), and ExcelAnnotator. We also demonstrate that pre-training TableParser on weakly annotated data allows highly accurate parsing of structured data in real-world table-form data documents. Fine-tuning the pretrained TableParser in various domains has shown large improvements in detection accuracy. We have observed that the state-of-the-art for table extraction is shifting towards DL-based approaches. However, devising suitable tools to facilitate training of such DL approaches for the research community is still lacking. Hence, we provide a pipeline and open-source code and data to invite the active contribution of the community.
dation (Project Number 100018_204647) for supporting this research project. Ce Zhang and the DS3Lab gratefully acknowledge the support from Swiss National Science Foundation (Project Number 200021_184628, and 197485), Innosuisse/SNF BRIDGE Discovery (Project Number 40B2-0_187132), European Union Horizon 2020 Research and Innovation Programme (DAPHNE, 957407), Botnar Research Centre for Child Health, Swiss Data Science Center, Alibaba, Cisco, eBay, Google Focused Research Awards, Kuaishou Inc., Oracle Labs, Zurich Insurance, and the Department of Computer Science at ETH Zurich. Besides, this work would not be possible without our student assistants: We thank Ms. Ada Langenfeld for assisting us in finding the Hungarian scans and annotating the tables; we thank Mr. Livio Kaiser for building an ExcelAnnotator prototype during his master thesis. We also appreciate the users’ insights on LayoutParser [24] shared by Mr. Cheongyeon Won. Moreover, the comments and feedback from Sascha Becker and his colleagues at SoDa Labs, Monash University, are valuable in producing the current version of TableParser. We also thank Sascha and Won for providing us with various South Korean/European table scans. Finally, we thank the reviewers at SDU@AAAI22 for carefully evaluating our manuscripts and their constructive comments. [9] W. Farrukh, A. Foncubierta-Rodriguez, A.-N. Ciub- [22] Y. Wu, A. Kirillov, F. Massa, W.-Y. Lo, R. Girshick, otaru, G. Jaume, C. Bejas, O. Goksel, M. Gabrani, Detectron2, https://github.com/facebookresearch/ Interpreting data from scanned tables, in: 2017 detectron2, 2019. 14th IAPR International Conference on Document [23] J. Eberius, C. Werner, M. Thiele, K. Braunschweig, Analysis and Recognition (ICDAR), volume 2, IEEE, L. Dannecker, W. Lehner, Deexcelerator: a frame2017, pp. 5–6. work for extracting relational data from partially [10] Y. Wang, I. T. Phillips, R. M. Haralick, Table struc- structured documents, in: 22nd ACM Internature understanding and its performance evaluation, tional Conference on Information and Knowledge Pattern recognition 37 (2004) 1479–1497. Management, CIKM’13, San Francisco, CA, USA, [11] S. R. Qasim, H. Mahmood, F. Shafait, Rethink- October 27 - November 1, 2013, He, Qi; Iyengar, ing table recognition using graph neural networks, Arun; Nejdl, Wolfgang; Pei, Jian & Rastogi, Rain: 2019 International Conference on Document jeev, 2013, pp. 2477–2480. URL: http://doi.acm.org/ Analysis and Recognition (ICDAR), IEEE, 2019, pp. 10.1145/2505515.2508210.
142–147. [24] Z. Shen, R. Zhang, M. Dell, B. C. G. Lee, J. Carlson, [12] J. Long, E. Shelhamer, T. Darrell, Fully convolu- W. Li, Layoutparser: A unified toolkit for deep tional networks for semantic segmentation, in: Pro- learning based document image analysis, arXiv ceedings of the IEEE conference on computer vision preprint arXiv:2103.15348 (2021).
and pattern recognition, 2015, pp. 3431–3440. [25] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learn[13] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: ing for image recognition, in: Proceedings of the Towards real-time object detection with region pro- IEEE conference on computer vision and pattern posal networks, Advances in neural information recognition, 2016, pp. 770–778.
processing systems 28 (2015) 91–99. [26] N. X. R. Wang, D. Burdick, Y. Li, Tablelab: An [14] O. Russakovsky, J. Deng, H. Su, J. Krause, interactive table extraction system with adaptive S. Satheesh, S. Ma, Z. Huang, A. Karpathy, deep learning, in: 26th International Conference A. Khosla, M. Bernstein, et al., Imagenet large scale on Intelligent User Interfaces, 2021, pp. 87–89. visual recognition challenge, International journal [27] Z. Chi, H. Huang, H.-D. Xu, H. Yu, W. Yin, X.of computer vision 115 (2015) 211–252. L. Mao, Complicated table structure recognition, [15] M. Everingham, L. Van Gool, C. K. Williams, arXiv preprint arXiv:1908.04729 (2019).
J. Winn, A. Zisserman, The pascal visual object [28] Y. Li, Y. Huang, Z. Zhu, L. Pan, Y. Huang, L. Du, classes (voc) challenge, International journal of Z. Tang, L. Gao, Rethinking table structure recogcomputer vision 88 (2010) 303–338. nition using sequence labeling methods, in: Inter[16] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Gir- national Conference on Document Analysis and shick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, Recognition, Springer, 2021, pp. 541–553. P. Dollár, Microsoft coco: Common objects in con- [29] W. Xue, B. Yu, W. Wang, D. Tao, Q. Li, Tgrnet: A text, 2015. arXiv:1405.0312. table graph reconstruction network for table struc[17] S. Raja, A. Mondal, C. Jawahar, Table structure ture recognition, arXiv preprint arXiv:2106.10598 recognition using top-down and bottom-up cues, (2021). in: European Conference on Computer Vision, [30] R. Long, W. Wang, N. Xue, F. Gao, Z. Yang, Y. Wang, Springer, 2020, pp. 70–86. G.-S. Xia, Parsing table structures in the wild, in: [18] X. Zheng, D. Burdick, L. Popa, X. Zhong, N. X. R. Proceedings of the IEEE/CVF International ConferWang, Global table extractor (gte): A framework ence on Computer Vision, 2021, pp. 944–952. for joint table identification and cell structure recog- [31] D. Nazir, K. A. Hashmi, A. Pagani, M. Liwicki, nition using visual context, in: Proceedings of the D. Stricker, M. Z. Afzal, Hybridtabnet: Towards IEEE/CVF Winter Conference on Applications of better table detection in scanned document images, Computer Vision, 2021, pp. 697–706. Applied Sciences 11 (2021) 8396. [19] J. Jiang, M. Simsek, B. Kantarci, S. Khan, Tabcellnet: [32] S. Luo, M. Wu, Y. Gong, W. Zhou, J. Poon, Deep Deep learning-based tabular cell structure detec- structured feature networks for table detection and tion, Neurocomputing 440 (2021) 12–23. tabular data extraction from scanned financial doc[20] M. Li, L. Cui, S. Huang, F. Wei, M. Zhou, Z. Li, ument images, arXiv preprint arXiv:2102.10287 Tablebank: A benchmark dataset for table detection (2021).
and recognition, 2019. arXiv:1903.01949. [21] K. He, G. Gkioxari, P. Dollár, R. Girshick, Mask r-cnn, in: 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980–2988.
doi:10.1109/ICCV.2017.322.
RPN Candidate Regions Input Image ResNet 101 FPN Feature Maps
plete TableParser pipeline have been made available at https://github.com/DS3Lab/TableParser.
The 10-minute lightning presentation at SDU@AAAI22 to the paper could be found under this recording, starting at 1:42:35.
As we have discussed in Section 5.2, we show the qualitative outputs of ModernTableParser and HistoricalTableParser on various types of inputs in Figures 7, 8, 9, and 10.14 The quality of structure parsing varies across inputs, but overall, the quality is high.
14Input and annotated figures of original size can be found under https://github.com/DS3Lab/TableParser/tree/main/figures. RoI Align
Fully Connected
Layers Convolution
Layers
Bounding Box Regression
Class Prediction Mask Prediction