Datasets:

kotmayyaka
/

full_page_hwr_rus

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Mixed COCO Handwritten Dataset

A dataset of handwritten notebook/page images with text-region polygon annotations and associated text fields.

The dataset is built as a mix of multiple public sources plus internally generated synthetic data.

It also includes model inference results produced by several handwriting detection/OCR models.

Dataset composition

This dataset is a combination of:

Internally generated synthetic data
ai-forever/school_notebooks_RU
ai-forever/school_notebooks_EN
AntiplagiatCompany/HWR200

Models used for inference

Predictions for some images were produced with the following models:

kotmayyaka/hwr_text_detection_rus — handwritten text detection (line/region polygons)
kotmayyaka/hwr_text_ocr_rus — OCR (text recognition)
kotmayyaka/hwr_text_kelm_rus — OCR / language-aware postprocessing (kelm)

Annotation format

Label

The dataset provides a prediction file in a per-image dictionary format:

Example structure:

{
  "94_11.jpg": {
    "predictions": [
      {
        "polygon": [[861, 1904], [860, 1905], [856, 1905], [855, 1906]],
        "text": ""
      }
    ]
  }
}

Where:

Top-level keys are image file names.
Each image has a predictions list.
Each prediction may include:
- polygon: polygon as a list of [x, y]
- text: recognized text

Intended use

Handwritten text detection (line/region segmentation)
Handwritten text recognition (OCR) for RU/EN
End-to-end handwritten page reading (detection + OCR)
Semi-supervised / pseudo-label training using model predictions