samiuc/pixparse-idl

IDL-WDS OCR Evaluation Dataset

Dataset Description

This dataset is a carefully curated subset of the original pixparse/idl-wds dataset, specifically designed for OCR evaluation and benchmarking.

Dataset Summary

• Source Dataset: pixparse/idl-wds - Industry Documents Library (IDL)
• Purpose: OCR evaluation on single-page documents
• Sample Count: 1,000 carefully selected single-page documents
• Selection Criteria: Only documents with exactly 1 page in their JSON metadata
• Format: Organized folder structure with paired image and ground truth data

Key Features

• Single-Page Focus: All documents contain exactly one page, eliminating multi-page complexity for OCR evaluation
• High-Quality Ground Truth: Each sample includes detailed OCR annotations with bounding boxes, polygons, and confidence scores
• Standardized Format: Consistent file structure across all samples
• Ready for Evaluation: Pre-processed and organized for immediate use in OCR benchmarking

Dataset Structure

File Organization

Each sample is stored in its own folder named by the document key:

document_key_1/
├── image.tif          # Document image in TIFF format
└── data.json          # OCR ground truth annotations
document_key_2/
├── image.tif
└── data.json

Data Format

Image Files (image.tif)

• Format: TIFF (Tagged Image File Format)
• Content: Single-page document images
• Source: Original document pages from the IDL collection

Ground Truth Files (data.json)

The JSON schema follows the original IDL-WDS format:

{
  "pages": [
    {
      "text": [
        "Line 1 of text",
        "Line 2 of text",
        "..."
      ],
      "bbox": [
        [left, top, width, height],
        [left, top, width, height],
        "..."
      ],
      "poly": [
        [
          {"X": x1, "Y": y1}, 
          {"X": x2, "Y": y2}, 
          {"X": x3, "Y": y3}, 
          {"X": x4, "Y": y4}
        ],
        "..."
      ],
      "score": [
        confidence_score_1,
        confidence_score_2,
        "..."
      ]
    }
  ]
}

Schema Details

• text: Array of text lines in reading order
• bbox: Bounding boxes in [left, top, width, height] format (normalized coordinates 0-1)
• poly: Polygon coordinates for each text line (4 corner points)
• score: Confidence scores from Amazon Textract OCR (0-1 range)
• Coordinates: All spatial coordinates are normalized relative to page dimensions

Usage

Loading the Dataset

import json
import os
from PIL import Image

def load_sample(sample_folder):
    """Load a single sample from the dataset"""
    image_path = os.path.join(sample_folder, "image.tif")
    json_path = os.path.join(sample_folder, "data.json")
    
    # Load image
    image = Image.open(image_path)
    
    # Load ground truth
    with open(json_path, 'r', encoding='utf-8') as f:
        ground_truth = json.load(f)
    
    return image, ground_truth

# Example usage
base_dir = "idl_wds_extracted"
sample_folders = [f for f in os.listdir(base_dir) 
                  if os.path.isdir(os.path.join(base_dir, f))]

# Load first sample
image, gt = load_sample(os.path.join(base_dir, sample_folders[0]))
print(f"Image size: {image.size}")
print(f"Number of text lines: {len(gt['pages'][0]['text'])}")

Dataset Statistics

• Total Samples: 1,000 single-page documents
• Source Documents: Filtered from ~19M pages in original IDL dataset
• Document Types: Legal documents, internal communications, reports, and other industry documents
• Text Languages: Primarily English
• Time Period: Historical industry documents (various decades)

Licensing and Usage

This dataset inherits the licensing terms from the original IDL dataset:

• License: IDL-train license (see original dataset for full terms)
• Attribution: Please cite the original IDL and IDL-WDS datasets

Citation

If you use this dataset, please cite the original work:

@dataset{idl_wds_2023,
  title={Industry Documents Library - WebDataset Format},
  author={Pablo Montalvo and Ross Wightman},
  url={https://huggingface.co/datasets/pixparse/idl-wds},
  year={2023}
}

Quality and Characteristics

Selection Process

• Documents were filtered to include only those with exactly 1 page
• Multi-page documents were excluded to ensure consistency
• All samples verified to have both image and JSON ground truth data

Ground Truth Quality

• OCR annotations generated using Amazon Textract
• Confidence scores provided for quality assessment
• Reading order preserved through columnar detection heuristics
• Bounding boxes and polygons for spatial understanding

Recommended Use Cases

• OCR model evaluation and benchmarking
• Text detection algorithm testing
• Document layout analysis research
• Reading order evaluation
• OCR confidence score analysis

Data Limitations

• Historical Bias: Documents reflect historical industry perspectives
• OCR Quality: Ground truth quality depends on Amazon Textract performance
• Document Variety: Limited to industry document types from IDL collection
• Single Page Only: Multi-page document scenarios not covered
• Language: Primarily English language documents

Contact and Support

• Original Dataset: pixparse/idl-wds
• IDL Contact: Kate Tasker, UCSF (kate.tasker@ucsf.edu)
• Technical Contact: Pablo Montalvo (pablo@huggingface.co)

For questions about this specific subset, please refer to the original dataset maintainers.