The provided dataset is composed of thousands of images captured through different cell phone cameras in which documents are present in various real-world backgrounds, with non-uniform patterns, consisting of different textures, colors, lighting, and image resolution. The images are in RGB Color and .png format. The ground truth of this dataset, which is described below, is different for each challenge.
For the first challenge, the interest region is the document zone, for the second one the text zone, and for the third one the interest regions are the handwritten signatures pixels. For all challenges, interest regions are represented by white pixel areas, while non-interest regions are represented by black pixel areas.
The documents contain personal information that cannot be made public. Therefore, we replaced the original data in the document with synthetic data. Handwritten signatures were acquired from the MCYT and GPDS databases, which were synthetically incorporated into the images of the identification documents. The database will also include ground truth images and will be used for training models and for calculating evaluation metrics.
To evaluate the methods, the following similarity metrics will be considered:
- Dice Similarity Coefficient (DSC);
- Scale Invariant Feature Transform (SIFT);
- Structural Similarity Index (SSIM).