Skip to main navigationSkip to main content
The University of Southampton
The Hummingbird Project

Research Round-Up

Published: 17 August 2018

A brief summary of a few recent articles that caught our eye. This month is all about the face, and how computer algorithms overcome the challenge of facial recognition.

Computer algorithms aim to generate a 3D face from a 2D image

As humans, we take for granted the ability to seamlessly extract 3D information from a 2D image. This ability to extract 3D information is critical for human facial recognition. In recent years, computer scientists have developed algorithms that can generate a 3D face from a single 2D image, a process referred to as 3D facial reconstruction. However, the 3D ‘faces’ reconstructed by these algorithms are not always the most accurate. In a recent paper, Feng and colleagues organised a competition to provide a benchmark dataset of facial images against which the accuracy of 3D facial reconstruction algorithms could be tested. The team then tested three 3D face reconstruction algorithms on facial images that varied in viewpoint and lighting. Their results showed highest performance for algorithms using texture versus landmark-based information. These findings provide an important benchmark for evaluating the accuracy of 3D face reconstruction algorithms and may help to improve the accuracy of future biometric algorithms.

Feng, Z. H., Huber, P., Kittler, J., Hancock, P., Wu, X. J., Zhao, Q., ... & Rätsch, M. (2018, May). Evaluation of Dense 3D Reconstruction from 2D Face Images in the Wild. In Automatic Face & Gesture Recognition (FG 2018), 2018 13th IEEE International Conference on (pp. 780-786). IEEE.


Facial recognition systems based on the primate visual system

Humans are remarkably good at recognising faces, and even have specific brain regions that specialise in the processing of faces. Based on the structure of the primate visual system, a certain class of algorithms called deep convolutional neural networks (DCNNs). The main benefit of these algorithms is their relative accuracy in recognising faces across variations in viewpoint, expression, and lighting. However, the means by which they actually represent faces (their “face code”) is poorly understood and lack explanability. In a new paper, O’Toole and colleagues review what is currently known about how these algorithms code and identify faces in relation to current psychological theories of human face recognition.

O’Toole, A. J., Castillo, C. D., Parde, C. J., Hill, M. Q., & Chellappa, R. (2018). Face Space Representations in Deep Convolutional Neural Networks. Trends in Cognitive Sciences.


There’s Wally!

And just for a bit of fun, a new robot equipped with facial recognition technology find Wally (or Waldo for the Americans) in 4.45 seconds!


Privacy Settings