A Distraction Score for Watermarks

In this work we propose a novel technique to quantify how distracting watermarks are on an image. We begin with watermark detection using a two-tower CNN model composed of a binary classification task and a semantic segmentation prediction. With this model, we demonstrate significant improvement in image precision while maintaining per-pixel accuracy, especially for our real-world dataset with sparse positive examples. We fit a nonlinear function to represent detected watermarks by a single score correlated with human perception based on their size, location, and visual obstructiveness. Finally, we validate our method in an image ranking setup, which is the main application of our watermark scoring algorithm.