sciencetech-blog

comments

Our brains are hardwired to recognise faces, even in inanimate objects, but researchers wanted to reveal how we can become virtually unrecognisable simply by tweaking our features.

Chicago-based Robert Woodley and Adelheid Mers used software to identify 60 facial dimensions before plotting them across more than 1,000 photos donated by people across the globe.

From the range of images, the team constructed an average face, and then used the dimensions to create what they call an 'Anti Face' for each contributor.

Developers Robert Woodley and Adelheid Mers used software to identify 60 facial dimensions, which were plotted across more than 1,000 photos donated by people from across the globe. For each photo, the team constructed an average face (pictured), before using the dimensions to create what they call an Anti Face 

The Anti Face is the exact opposite of the face being studied, in terms of shape, size and individual features, said the researchers. 

The relevant 'opposite' feature is taken from all the photos submitted, and sits at the other end of the scale.

The Synthetic Face website, created by Mr Woodley and Mr Mers, reveals how faces can be altered by using sliders.

To plot the dimensions, the researchers used Eigenfaces analysis - a statistical face recognition technique.

It uses Principal Component Analysis to calculate a set of Eigenvectors, or Eigenfaces.

These Eigenfaces can be thought of as 'face ingredients'.

To calibrate the model, the researchers calculated 60 'ingredients' of points taken from more than a thousand faces.

Each time a new face was uploaded, it was compared to these points using a linear combination of the individual Eigenfaces.

According to the research, the average face (pictured) is androgynous. It is not race-neutral because the original sample set was biased towards caucasian faces. The researchers said they hope to fix this problem in the future as they collect more faces

To plot the dimensions, the researchers used Eigenfaces analysis - a statistical face recognition technique. These Eigenfaces can be thought of as 'face ingredients'. To calibrate the model, the researchers calculated 60 'ingredients' taken from random faces (pictured)

Initially, the researchers explored the concept of face continuum, or how unique a person's face is. 

'From the face continuum emerged the notion of the Anti Face,' said Mr Woodley.

'The Anti Face program is face recognition turned upside-down.

'After looking at your image, it creates a face as different from yours as possible. It might change your gender, your age, your expression.'

This means prominent features in the reconstructed faces, with high weighting, would be equally prominent, though opposite, in the anti face.

Each time a new face is uploaded, it is compared to these points and is created using a linear combination of the individual Eigenfaces. Sliders along the side of each image (pictured) can be used to change the features across each of these 60 points to see the face change from male to female, for example

'The Anti Face program is face recognition turned upside-down.' said the researchers. 'After looking at your image, it creates a face as different from yours as possible.' This means prominent features (highlighted in shadows pictured) would be equally prominent, though opposite, in the anti face

'The terms Face Detection and Face Recognition are often used interchangeably but to be precise, Face Detection picks out faces from an image, whereas Face Recognition takes a new image of a face and matches it with one it has previously seen,' continued Mr Woodley.

'Our research into multi-dimensional visual spaces draws upon algorithms used in Face Recognition.'

'[With the average face], all the coordinates are zero. This face resides exactly in the middle of the face cloud.'

According to the research, the average face is androgynous. It is not race-neutral because the original sample set was biased towards caucasian faces.

The researchers said they hope to fix this problem in the future as they collect more faces.

Photos can be submitted via Robert Woodley's Devart project page.