I stepped out to buy a loaf of bread on Saturday and found a parking apron taken up by a trio busy stringing marigolds into streamers. Of course, the time-keeper in my head told me, “Tomorrow is dusshera.” It wasn’t very easy to track this year, with no lights, no festivities, no night-long dances. But it was hard to forget too, with all the sighing and complaining from friends and family about missing everything this year. I didn’t actually, I’m a happy bear these days. I’m a day or two late, but a happy dusshera to you.
But while lost in uffish contemplation of these photos snapped off quickly with my phone, a monstrous jabberwock raised its head. Do you recognize a subtle bit of structural racism here? Let me explain. A phone camera is no longer the simple light catching device that a camera was forty years ago. The image is mediated by software, and the huge improvement in this decade has been due to the addition of AI to the mix. That tweaks the exposure and colour according to optimizations it has been trained to use. Many of these improvements involve recognizing and enhancing faces. Just check your photos from five years back, and you’ll see how much better faces look in recent photos. I say usually, because it depends on the skin colour.
The AI is trained on data sets selected by Google, and these are largely biased to lighter skin colours. See, for example, a photo taken three years ago, by a six year old phone, in Madrid’s Atocha station. I have taken a random face from the shadows and inset it to show how well the AI has captured that. All I did was to increase the overall illumination in the inset. See the subtlety with which the details and shadows are rendered. Compare that to the inset face (with the same tweak for overall illumination) visible in the photo above that: the AI has rendered the bamboo framework in the background better than the face. This comparison shows that it is not light or darkness that makes a difference; it is algorithmic bias. It took me a bit of work on the featured photo to make the face visible. It is because of this bias that a DSLR or a good bridge camera still remains superior to a smart phone for street photography in India and most of the world.