It might even be that Google Research has found a way to stop you from looking silly in group photos. We all know that taking a group photo is hard. Just try to keep everyone's attention while you compose the shot! And you can be sure that in any group photo at least one person has their eyes shut at the moment you hit the button.

The solution to the problem is computational photography. Take more than one group picture and use the set to create one good photo. Sounds easy but getting it right needs a lot of attention to detail. This is what Rajvi Shah and Vivek Kwatra at Google Research have been working on. 

Given a set of group pictures the first task is to score each one by applying face detection and then working out a score based on how many of the faces are oriented correctly, smiling and have their eyes open. Using this, the images can be ranked and the best, but possibly still flawed, group picture can be selected. 

The next stage goes back to the set of rejected pictures and attempts to find any faces that have high scores that could be used to replace faces that have low score in the best photo. 

The real difficulties here are in selecting faces that are actually better then the original and a lot of work has been put into selection criteria. Learning algorithms were used to tell the difference between smiling and non-smiling and eyes open and close faces. 

What can one say about the results?

As the authors of the paper say:

We have demonstrated the effectiveness of our approach through a variety of examples that bring a smile on people’s faces.

When you look at the before and after photos there is a slight spookiness in imagining a world where everyone smiles all of the time and perhaps the loss of the amusing and memorable group photo might be something to mourn but... do you want to be the one with your eyes closed or looking like a wet weekend?

The final comment gives you an indication of where things might be going:

In future, we would like to optimize our implementation for real-time performance, making it an attractive utility for computational cameras. We would also like to learn and incorporate more subtle facial attributes for goodness evaluation.