[AxesPushPatty]

The problem is more in the case that if you have access to the ground truth data, and use such a tool to create images without detectable faces, you can artificially create more testing data. In the end this training data will be more robust, and the underlying method, like any form of CNN, will be able to generalize better.