Mooney Face Classification and Prediction

Tsung-Wei Ke, Stella X. Yu, David Whitney

UC Berkeley / ICSI

Mooney faces are special two-tone images from which human observers can effortlessly identify many attributes of the person. How such rich perception is achieved upon such poor information could shed light into face recognition in general. While Mooney faces are important, there are only a small number of instances hand-crafted from source photos which are no longer available. Our first goal is to generate a plausible Mooney face automatically from any face photo. We are then able to create a large-scale face dataset with paired grayscale and two-tone images. Our second goal is to predict a fine-tone grayscale version from a two-tone Mooney face. We build GoogleNet and conditional Generative Adversarial Network models to learn Mooney face classification and prediction respectively. Our predicted faces from never seen Mooney images striking resemblance to source photos, demonstrating great potentials of combining deep learning and Mooney faces for more effective face recognition.


Berkeley Mooney Dataset v1.0

Tsung-Wei Ke, Stella X. Yu, and David Whiteney. "Mooney Face Classification and Prediction by Learning across Tone", in ICIP2017

Tsung-Wei Ke, Stella X. Yu, and David Whiteney. "Mooney Faces from Photos", in VSS2017


  • Our mooney face classifier is fine-tuned from openface.
  • To generate Berkeley Mooney Dataset v1.0, we use the cropped and aligned Facescrub Dataset provided by MegaFace (by downloading this dataset, you agree to the conditions of the creators of the data)
  • Our ground-truth mooney faces are provided by David Whitney and PICS