README.md

metadata

license: cc-by-4.0
task_categories:
  - visual-question-answering
  - image-to-text
tags:
  - optical-character-recognition
  - visual-question-answering
  - robustness
  - watermark
  - numerical
  - multimodal
pretty_name: FADE
size_categories:
  - 1K<n<10K

FADE Watermark OCR Dataset

Overview

This dataset contains watermarked images, their corresponding masks, the alpha values used for watermarking, and the actual text embedded (as a 9-digit number). It is designed to train and evaluate OCR models in the presence of watermarks.

Paper: FADE: Probing the Limits of VLMs on fine-grained OCR

Data Fields

Column Name	Data Type	Description
Image with watermark	binary	The raw binary bytes of the watermarked image (JPEG). When loaded with Hugging Face datasets, this can be cast to a PIL Image.
Masked image	binary	The raw binary bytes of the corresponding mask image.
alpha value	float	The blending factor used to apply the watermark (e.g., 0.05, 0.1, 0.5, 1.0).
Actual answer	string	The ground truth 9-digit watermark text embedded in the image.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("deep9539/FADE-watermark-ocr")

# View the structure
print(dataset)