So, when I read that Google has applied machine learning to tag LIFE’s 4-million photographic archives, it seemed like a great idea. But what exactly is machine learning? Just as the term “AI” used to be thrown around in the past, nowadays it’s the term “machine learning” that gets brandied about and picked up and applied to everything in sight.
Is Google’s LIFE TAGS program really learning, getting better as it goes along, making mistakes and learning from them? Or, is it just an image recognition algorithm that can figure out that, oh there’s a hat in this image and so let me tag it as a hat, etc?
I decided to give it a try. Type in camera and you get a list of key phrases with camera in it. I select “mirrorless interchangeable-lens cameras” and would you believe it listed images of cameras that it think are mirrorless, plus even gave a very accurate description of what a mirrorless camera is, including the why of the term “mirrorless,” but stumbling just a bit on the auto-focus part. (I am sure the text description has not been “learned” by scouring the Internet but has been typed in by a carbon-based life form.)
I’m impressed that Google accurately described what mirrorless cameras are. There are a lot of misconceptions out there, including the wrong belief that every camera that lacks a mirror is technically mirrorless.
However, the author was just a bit optimistic about the mirrorless C-AF capability, as it stands currently. The statement “MILCs use high speed contrast detection auto-focus, rather than the phase-based auto-focus used in SLR cameras.” is contradicted later by: “The latest generation of mirrorless cameras, however, have PDAF pixels built into the image sensor, offering fully competitive and accurate auto-focus that are many times faster during continuous shooting with continuous auto-focus than DSLRs.”
The last statement about mirrorless continuous shooting C-AF being “many times faster” than DSLR continuous shooting C-AF is just a bit ahead of its time, when comparing the best flagship pro mirrorless and DSLRs as far as being able to consistently get spot-on in-focus shots using Lock-on subject tracking. Though a mirrorless can shoot faster (continuous shooting), it’s the consistent in-focus subject tracking (tracking C-AF) part that is still the one small bragging right of pro DSLRs. It’s coming, but we’re not quite there yet. Give it a couple of years.
But there’s no doubt that mirrorless “can be considered a further evolution phase of the conversion of an SLR camera into a digital system.”
The biggest problem, however, is that the cameras depicted are simply not mirrorless. “Learning” implies a method to tell the program it erred and needs to correct its wrong assumptions. Unfortunately, there is no such functionality, so is this really “machine learning?” As it stands now, Google’s LIFE TAGS experiment is at best a faulty image recognition algorithm, as far as being able to accurately recognize and tag mirrorless interchangeable-lens cameras. But, if anybody can do it, Google engineers can. Give it a couple of years.
To be fair to Google, are there any mirrorless cameras to be found in the LIFE archive photos? Mirrorless cameras are a relatively recent development (the Panasonic G1, the first mirrorless camera, came out on September 12, 2008) and LIFE stopped publishing way before that (the last weekly issue was published on December 29, 1972). Granted, there have been intermittent publication under the LIFE brand after 1972, but would any pro photographers have used mirrorless cameras back then, and were mirrorless cameras popular enough as to figure in many of the pictures taken by LIFE pro photographers and make it into the publication?
The “mirrorless interchangeable-lens cameras” algorithm should probably be tweaked so that only cameras from known mirrorless camera manufacturers should be considered, and then only those from photos dating after September 12, 2008. That might perhaps give a slightly more accurate result.
Visit our Featured Site: Google Arts & Culture Experiments – LIFE TAGS