Seeing Faces
When you look at this image, what do you see?
A lot of people will see faces before they see anything else. And you could wonder ... if humans have this ability, what about computer vision algorithms? It turns out that some researchers were wondering the same thing and this led to a new dataset as well as this interesting paper.
The dataset contains annotations of places where faces are located. They also keep track of some extra things like how hard it was to find that face, as well as gender. With this dataset, the authors wondered if state of the art face detectors were able to detect these "faces" and if finetuning might improve their results. It turns out that they were able to get better performance for this task by finetuning a state of the art face detector on a dataset with animal faces!
The paper concludes by also offering some thoughts on what circumstances folks might start seeing faces in random objects. It turns out that random noise can trigger the face detector but only if you get the noise just right. Too granular and it just looks like white noise, too smooth and it just looks like a blob. But somewhere in the middle lies a sweet spot.
The paper has lots of interesting details, and if you're keen on getting the short story I recommend checking the website for the paper. It contains a short video that gives a quick overview of the work.