Automatically keyword numbers in photos?

[atolkachev]

Lightroom Version Number

Lightroom Classic 11.3.1

Operating System

1. macOS 12 Monterey
2. iOS

Is there a plugin or a hack to either automatically keyword or at least search (that would probably apply more to LR) by the numbers visible in the photos?

This is helpful to sports photographers. Not every shot will show a jersey number clearly, but if a half does and get thus indexed, that could be hours' worth of work savings on a sizable shoot.

Thanks!

 

[Califdan]

See https://johnrellis.com/lightroom/anyvision.htm

scroll down in the article and John mentions some other tools as well.

 

[atolkachev]

I read that and might have added my question to that thread. What else is there? Excire may or may not be dead. Their online demo is down, and their forum is spammer playground. None of the other tools seem practicable. John's own plugin may be the only tool available, but with the Google dependency and the per image (or per 1000 images) fee I just can't get too excited about it.

LR cloud search tries, but its results are widely inaccurate and incomplete so is to make it unusable.

*sigh*

Come to think of it, relying on any sort of AI to partially index any given sporting event may be fool's gold. Let's say there is a series of action shots of a player, where the number is visible in half the shots and partially or fully obscured in the rest. A good AI will tag half the images leaving it to the human to tag the rest. But what does the human have to go on if not the neighboring shots (which have already been tagged and perhaps excluded from the secondary tagging exercise)? So now they need to be unexcluded and looked at. And so the whole series might have been tagged by a human more effectively in one shot.

 

[johnrellis]

There's another service that was recently upgraded, Sportvisionai, that I haven't added to the list on the Any Vision page yet:
https://community.adobe.com/t5/ligh...n-for-sports-photography/m-p/12918143#M272965

It's brand new, and I haven't received any feedback on how well it works or compares to Google AI. I'd be only mildly surprised it were better than Google AI, which has made very large improvements in number OCR in the years since I first released Any Vision.

 

[johnrellis]

Come to think of it, relying on any sort of AI to partially index any given sporting event may be fool's gold. Let's say there is a series of action shots of a player, where the number is visible in half the shots and partially or fully obscured in the rest. A good AI will tag half the images leaving it to the human to tag the rest. But what does the human have to go on if not the neighboring shots (which have already been tagged and perhaps excluded from the secondary tagging exercise)? So now they need to be unexcluded and looked at. And so the whole series might have been tagged by a human more effectively in one shot.

I think it depends on the event and the kind of photography. I've got a fair number of customers who do event photography where they set up at a fixed spot and get photos from an angle where they get clear, well-lit views of the numbers on runners, cyclists, dirt bikes. They get pretty good accuracy (*), and they find Google's price of $1.50 / 1000 photos economically worthwhile (the first 1000 photos each month are free).

At least some of the photographers don't review the OCR results -- prior to using OCR, their customers would scroll through dozens or hundreds of unlabeled photos (based on event and timestamp) to find pics of themselves, which they would buy. So even if the OCR is mediocre, it means that customers are more likely to find their pics quickly.

I think that number OCR for sports like basketball and football would be much less useful, since even if the OCR were perfect, the numbers would be obscured in a large fraction of the photos.

(*) For these kinds of events, e.g. running races, the percent of numbers extracted from photos (recall) is pretty good, and the percent of the extracted numbers that are correctly recognized (precision) also pretty good. One photographer who did running races estimated it at above 85 - 90%.

 

[atolkachev]

I think it depends on the event and the kind of photography. I've got a fair number of customers who do event photography where they set up at a fixed spot and get photos from an angle where they get clear, well-lit views of the numbers on runners, cyclists, dirt bikes. They get pretty good accuracy (*), and they find Google's price of $1.50 / 1000 photos economically worthwhile (the first 1000 photos each month are free).

At least some of the photographers don't review the OCR results -- prior to using OCR, their customers would scroll through dozens or hundreds of unlabeled photos (based on event and timestamp) to find pics of themselves, which they would buy. So even if the OCR is mediocre, it means that customers are more likely to find their pics quickly.

I think that number OCR for sports like basketball and football would be much less useful, since even if the OCR were perfect, the numbers would be obscured in a large fraction of the photos.

(*) For these kinds of events, e.g. running races, the percent of numbers extracted from photos (recall) is pretty good, and the percent of the extracted numbers that are correctly recognized (precision) also pretty good. One photographer who did running races estimated it at above 85 - 90%.

Running is about one sport I don't do. Football, soccer, lacrosse, basketball, volleyball. I wanted AI to tag what it can, they I tag the rest (facial recognition might play a part, too). But what I really need is low cost elves.

 

[NicoSportVisionAi]

There's another service that was recently upgraded, Sportvisionai, that I haven't added to the list on the Any Vision page yet:
https://community.adobe.com/t5/ligh...n-for-sports-photography/m-p/12918143#M272965

It's brand new, and I haven't received any feedback on how well it works or compares to Google AI. I'd be only mildly surprised it were better than Google AI, which has made very large improvements in number OCR in the years since I first released Any Vision.

Hi, I'm the owner of SportVisionAi
Actually it works better than Google OCR in some edge cases. Here are two examples where I compare the Google OCR vs SportVisionAI.

Here, Google fails to detect the 11 as it is partially hidden. My AI detects it:

Here, Google thinks it is "EE" instead of 93. My AI interprets it properly as 93:

If you have any questions, don't hesitate!