Is key wording becoming a thing of the past

Not quite yet but image content recognition algorithms are getting much, much better:

http://www.nytimes.com/2014/11/18/science/researchers-announce-breakthrough-in-content-recognition-software.html?

hp&action=click&pgtype=Homepage&module=second-column-region&region=top-news&WT.nav=top-news

<p>Suits me. I'm looking forward to better content recognition tools. I'm already using Picasa's facial recognition tool to help organize my photos of people and tag them, which is much better than trying to find and tag them in Lightroom. The only downside is that tags/keywords are not consistently recognized for raw files. So I still need the JPEGs to associate the same raw files with the desired keywords and names to identify people in those photos.</p>

<p>This is yet another area where Google could make a Lightroom killer and serious challenge to Adobe, by integrating existing content recognition tools into an advanced version of Picasa (which still desperately needs better raw processing, noise reduction, white balance and other tools to be a viable challenger to Lightroom).</p>

<p>Would save a lot of time given the backlog of keywording I still have left to do!</p>

<p>From what I gather from the article (which isn't terribly clear on the actual implementation), this sounds more like a web/cloud API, rather than something that runs local on your PC. Google probably is more than able to turn this into a API that others (Adobe and its competitors) could implement, which would be nicer than them trying to write desktop software (which isn't their strength by any stretch of imagination). Concerns on data privacy would creep in, though, plus the usual anti-cloud crowd making noise.Yet, it would enable this technology for any program, and given that some companies have products that can compete with Lightroom, it would be nice to keep this a bit level playfield for competition (and innovation) sake.</p>

<blockquote>

<p>Would save a lot of time given the backlog of keywording I still have left to do!</p>

</blockquote>

<p>You too?<br>

<br />I gave up keywording my images a long time ago mainly because I never needed to hunt up a photo by trying to remember a keyword I assigned it whether I could remember it or not. Since no one would ask for a specific photo they saw in my online gallery I saw no need for it and if one of them did ask me, they'ld just provide the URL link to the one in question they saw online.</p>

<p>I don't think this helps me all that much because it doesn't recognize what is <em>important</em> to me about photos, only what is in them. For example, let's say someone asks me for a photo of someone with interesting tattoos. Probably 70% of the people I photograph have visible and extensive tattoos, and if they get correctly marked by a recognition program, I now have 70% of my people photos keyworded with "tattoo." However, only about 2% of those photos might be considered interesting for their tattoos, if that much. So instead of having a hundred photos to sort through, I have tens of thousands.<br>

<br />A similar example comes up with keywording automated to geotagging, which can already be done if a photo has location embedded. However, once again, that doesn't tell me if the location is something I care about. For 2014, I have some 20,000 photos taken <em>in</em> San Francisco. However, if I look at 2014 for my photos keyworded with San Francisco, which means that they are photos that are representative <em>of</em> San Francisco, I'm down to 300, a fairly easy number to sort through quickly.</p>

<p>So while all this automation makes certain things easier, such as finding photos of Lex's aunt Mathilda, it makes some types of selection far more difficult.</p>

<p>That's an impressive advance but is probably in its infancy akin to speech recognition in its early days, and presumably relies heavily on image quality to determine accuracy - humans can recognize a herd of elephants and count their numbers in a 75x75 pixel thumbnail which I imagine will be an incredible challenge for machine analysis. </p>

<p>I just read about an impressive real time machine vision system able to recognize a person in a crowd and follow him across multiple cameras with law enforcement application in mind - see the video in the page:<br>

<a href="http://www.washington.edu/news/2014/11/12/moving-cameras-talk-to-each-other-to-identify-track-pedestrians/">http://www.washington.edu/news/2014/11/12/moving-cameras-talk-to-each-other-to-identify-track-pedestrians/</a> </p>

<blockquote>

<p>"...finding photos of Lex's aunt Mathilda..."</p>

</blockquote>

<p>One does not find aunt Mathilda. Aunt Mathilda finds <em>you</em>.<br>

<br>

Ideally I'd like to see more content recognition with combined search "terms" comparable to text searches: aunt Mathilda in 2010 with Porkie the schnauzer; cousin Filbert with his tattoo reading "Sue" before he got divorced and had it changed to "Sue me if I ever tattoo another wife's name across my forehead."</p>

<p>People who sell rights to their online stock (me) will not be skipping key wording anytime soon.</p>

<blockquote>

<p>People who sell rights to their online stock (me) will not be skipping key wording anytime soon.</p>

</blockquote>

<p>How do future rights purchasers of your stock images first see your images, JH? By keywords? Or from viewing an online stock agency thumbnail?</p>