Skip to main content
Open Access Publications from the University of California

Predicting Facial Attractiveness Within an Individual’s Portraits using Siamese Networks

  • Author(s): Zhang, Angel Iek Hou
  • Advisor(s): Cottrell, Gary
  • et al.

People often have to decide which photograph of themselves to include in a portfolio or album, or share online. When making this choice, people are usually guided by what they think others will deem attractive. There is some consensus as to which portraits are the most attractive, and which ones are the least attractive. The goal of this thesis is to predict which photo of an individual is the most popular among viewers. In this work, we show experimentally that there is agreement among people regarding the attractiveness of an individuals photo and that it is possible to model and predict which photos are the most popular. We collected a dataset featuring portfolio pictures of actors and actresses, models and singers. Human votes for these images were also collected. We then trained a siamese network with a ranking loss function to predict the relative ranking of images. Predictions were made on the test set using one side of the network. We compared these predictions with the human votes and were able to achieve an accuracy of 61%.

Main Content
Current View