I admit it: I’m very impressed by self-processed photographs
This is a subject that fascinates me, and one I’ve discussed on several occasions, most recently in December of last year.
A couple of nights ago, I was in Porto, in northern Portugal. I arrived with my wife by car as night was falling, and decided to take a few snaps of the impressive view, including Vila Nova de Gaia on the opposite bank of the Douro river, lit up by neon signs advertising the locally produced port stored there in dozens of warehouses. When I’d finished I uploaded a couple of my pictures to my Instagram account, and forgot all about it.
Yesterday morning, Google informed me that I had an automatically created panoramic photo, shown above. My smartphone had detected I had several photos of the same area, it had selected a few and then managed to overlap them to create a wide-angle, nighttime cityscape. The feat is all the more impressive if you look at the individual photos used in the process: they’re not all perfectly horizontal, the color of the sky is different, the white balance doesn’t match, and of course the camera angles are all different. At no time when I was taking the snaps was I thinking of creating a panorama effect.
Needless to say, this isn’t about my photographic skills, but the software that created this image. This wasn’t done by Google minions or mechanical turks: everything took place automatically, but the result is something that until now we would have thought only humans would be capable of.
In other words, software with significantly superior criteria to a human has chosen the photos, adjusted them on the basis of color, tone, temperature, brightness, and made sure that each is joined seamlessly to the next. Believe me, if I tried to do this myself, it would take a lot longer, and the results would be inferior. If we enlarge the photo to its original size, we can see that the superimposition is perfect: you cannot see the join. But what is most impressive about all this the extent to which we have been able to create a program able to reflect human attributes such as artistic sensibility…
All of this in a piece of software: a photograph that is much better than any number out there in the world. And all I did was hold my smartphone up and take a few snaps. Everything else was done by machine. How many images have been used to train a machine learning algorithm to “see” what makes a photograph interesting, to be able to surprise us? The more I read and try to learn to keep up with my role as an strategic advisor at BigML, the more impressed I am.
Of course, my fascination is the result of me belonging to another age. I remember when these kinds of panoramic shots had to be created manually; I remember the first digital cameras, and I know how limited their possibilities initially were; I remember taking panoramic photos with a tripod, carefully trying to match up each section… so you’ll forgive me if I’m still impressed that a piece of software can do all this, and more, without any human input. Very impressed.
(En español, aquí)