September 2017
Volume 17, Issue 10
Open Access
Vision Sciences Society Annual Meeting Abstract  |   August 2017
Mooney Faces from Photos
Author Affiliations
  • Tsung-Wei Ke
    UC Berkeley
    International Computer Science Institute
  • Stella Yu
    UC Berkeley
    International Computer Science Institute
  • David Whitney
    UC Berkeley
Journal of Vision August 2017, Vol.17, 915. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Tsung-Wei Ke, Stella Yu, David Whitney; Mooney Faces from Photos. Journal of Vision 2017;17(10):915.

      Download citation file:

      © ARVO (1962-2015); The Authors (2016-present)

  • Supplements

Mooney faces are simply black and white pictures that human subjects can still effortlessly recognize as a face, along with many attributes such as age, gender, and facial expressions. How such a remarkable ability is achieved with so little information is key to understanding human visual recognition. However, a very basic question is what makes a Mooney face, as not every binarized face image appears as a Mooney face. Current Mooney face generation is ad-hoc, requiring manual grayscale image editing and human subjective judgment, a time and labor-consuming process. Our goal is to generate Mooney faces from photos automatically. We take a deep learning approach based on very limited Mooney faces but many more face and non-face photos. We first train a gray-scale face classifier using FaceScrub images (91,712 faces, 630 identities) as positive examples and random ImageNet samples (91,712 images, 1000 classes) as negative examples. We then train a black-white Mooney face classifier by fine-tuning the grayscale model on both thresholded face/nonface images and a small training set of Mooney face images. We assume that any non-face images, thresholded at any level, positive or negative, should be classified as non-faces, whereas face images should be classified as Mooney faces only in positive but not in negative contrast. Such instances are generated in balanced mini-batches to train our Mooney face classifier. Given a grayscale face image, we generate a set of two-tone versions by smoothing and thresholding the image with a range of parameters. The highest scoring image predicted by our Mooney face classifier becomes the Mooney face. If the highest score is below probability 0.5, we reject the Mooney face and declare the grayscale image unfit as a Mooney face candidate. This approach yields a large-scale Mooney face dataset, potentially useful for future computer vision, psychophysics, and neurophysiology experiments.

Meeting abstract presented at VSS 2017


This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.