CISOL

A HuggingFace mirror of the CISOL dataset (Zenodo record 10829550, CC-BY-4.0): German construction-industry steel ordering lists annotated for table detection and table structure recognition.

Original paper: Tschirschwitz et al., WACV 2025 (arXiv:2501.15469).

License

CC-BY-4.0. The peer-reviewed paper (WACV 2025) states: "The data is licensed under the Creative Commons Attribution 4.0 International license to support full research use, subject to proper anonymization during the preparation phase." Commercial use and redistribution are permitted with attribution.

Cite the original CISOL paper (see Citation below).

Configs

Config	Source	Images	Description
td_tsr	cisol_TD-TSR.zip	Full document pages	End-to-end table detection + structure recognition
tsr	cisol_TSR.zip	Pre-cropped table regions	Table structure recognition only

Splits

Split	Annotated	Available in
train	yes	td_tsr, tsr
validation	yes	td_tsr, tsr
unlabeled	no	td_tsr only

No public test split: the CISOL Zenodo release includes only train and val annotation JSON files; test annotations are withheld (standard competition holdback). The unlabeled split (2,436 images) contains full-page images from the same source corpus with no bounding-box annotations, useful for self-supervised or semi-supervised pre-training.

Annotation categories

COCO-format bounding boxes covering five structural elements:

category_name	Description
table	Full table bounding box (td_tsr only)
row	Individual row region
column	Individual column region
spanning_cell	Cell spanning multiple rows or columns
header	Header row or column region

Schema

Field	Type	Description
image_id	int32	COCO image ID (local to each split)
file_name	string	Original image filename
image	Image	Raw image bytes (PNG/JPEG)
width	int32	Image width in pixels (0 for unlabeled rows)
height	int32	Image height in pixels (0 for unlabeled rows)
origin_tag	string	Project-origin balancing tag (empty for unlabeled)
size_tag	string	Document-size balancing tag (empty for unlabeled)
type_tag	string	Document-type balancing tag (empty for unlabeled)
annotations	list[struct]	COCO annotations; empty list for unlabeled rows

Annotation struct fields: annotation_id, category_id (int32); category_name (string); bbox (list[float32], COCO [x, y, width, height]); area (float32); iscrowd (bool).

Segmentation masks (polygon/RLE) from the source COCO JSON are excluded; bounding-box coordinates are sufficient for TSR training.

Usage

from datasets import load_dataset

# Full-page images with detection + structure annotations
ds = load_dataset("rootsautomation/CISOL", "td_tsr")
row = ds["train"][0]
print(row["file_name"], len(row["annotations"]["bbox"]))

# Pre-cropped tables, TSR only
tsr = load_dataset("rootsautomation/CISOL", "tsr", split="train")

# Unlabeled images for pre-training
unlabeled = load_dataset("rootsautomation/CISOL", "td_tsr", split="unlabeled")

Citation

@inproceedings{tschirschwitz2025cisol,
  title     = {{CISOL}: An Open and Extensible Dataset for Table Structure
               Recognition in the Construction Industry},
  author    = {Tschirschwitz, David and Gekeler, Ella and Rodehorst, Volker},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications
               of Computer Vision (WACV)},
  year      = {2025},
}