wqzh/DocLayout-PageLocationDataset

English | 中文

Introduction

The original images of this dataset are from DocLayNet-v1.2. You can run src/1_save_img_from_parquet.py to convert the parquet files into images.

Page Number Cropping

This dataset uses the yolo11-medium-page-best-quant-ir9.onnx weights from Here to detect page number bounding boxes from the images obtained above, and then crop out the page number regions (detection errors may exist, as no post-processing was applied by the author).

The cropped images are saved in the images-crops directory, and the detection results are saved in the images-labels directory. The author added a constraint: the YOLO-format label file is only saved when exactly one page number box is detected, while crops of all detected boxes are retained. Therefore, the number of files in images-crops is greater than that in images-labels.

You can run src/2_det_onnx_wo_opencv_cuda.py to detect page numbers, crop images, and save label files.

find images-labels -type f -name "*.txt" | wc -l
find images-crops -type f -name "*.jpg" | wc -l

• Number of .txt files in images-labels: 15902
• Number of .jpg files in images-crops: 17203

Page Number Concatenation

src/3_concat_crops_by_pages.py concatenates all cropped images from images-crops and saves them in the images-crops-concat directory, so you can preview the entire dataset.

Example

Note: If you intend to use the label files from this dataset, they need to be cleaned separately.

Page Number Coordinate Distribution

src/4_visualize_yolo_txt_distribution.py visualizes the distribution of bounding box centers from all label files in images-labels.