bastienp/visible-watermark-pita

Visible watermarks datasets

We have observed that while datasets such as COCO are available for object detection, the availability of datasets specifically designed for the detection of watermarks added to images is significantly limited. Through our research, we identified only one such dataset, which originates from the paper Wdnet: Watermark-Decomposition Network for Visible Watermark Removal [1]. This dataset provides a collection of images along with their corresponding watermark masks for the purpose of watermark removal. Additionally, we noted that accessing this dataset presented challenges in terms of data accessibility and regeneration of dataset samples.

The CLWD Dataset, introduced in Wdnet: Watermark-Decomposition Network for Visible Watermark Removal [1], comprises images sourced from the COCO Dataset (Lin et al., 2014) [ 2] and masks of colored watermarks featuring random positions and opacities

Dataset Details (PITA Dataset)

We decided to introduce the Pita dataset, which is based on images from the COCO dataset (Lin et al., 2014) [ 2] and combines these with logos from the Open Logo Detection Challenge (Su et al., 2018) [3]. The dataset introduces several changes compared to other datasets, with a focus on the task of watermark detection rather than watermark removal.

The dataset is structured into three splits: a training split, a validation split, and a test split, collectively comprising approximately 20 000 watermarked images featuring both logos and text. We decided to incorporate two types of labels:

• Text: The images are watermarked with a random font available on the computer used for generation, and the text size is also randomized.
• Logos: The logos are sourced from the Open Logo Detection Challenge dataset (Su et al., 2018) and are characterized by random sizes and opacities.

The position of the logo or text is randomly selected from a set of available positions, specifically corners or the center. This restriction was introduced based on the observation that watermarks on social media or stock image websites are predominantly located in these positions.

The dataset is accompanied by command-line interface tools that facilitate reproducibility. These tools support both YOLO and Hugging Face formats, allowing the download of the dataset and generation with ease

Dataset Sources

• Repository: https://github.com/OrdinaryDev83/dnn-watermark
• Demo: https://huggingface.co/spaces/qfisch/watermark-detection

Uses

• DETR with Hugging face Transformers
• YoloV8 model with ultralytics
• FastRCNN with Pytorch Lighthning

Source Data

• COCO dataset (Lin et al., 2014) [ 2] and
• Open Logo Detection Challenge (Su et al., 2018) [3].

Data Collection and Processing

Generation of the dataset is reproducible using the cli tool of this repository.

A --help option is available in order to describe how to use the tool.

Annotation process

Logo were added to COCO images by applying rotation, scaling and opacity changes at a random position on the image. on the image.