I’m sure there was a thread on ST a while back linking to one of those marathon investigation stories about a women who habitually forged entry to Disney races (which I think are expensive and oversubscribed)
They did a load of digging to prove where/when she’d done it and actually arranged for the police to arrest her at the start of the next race. I think she got done for fraud.
ETA - Police confront suspected bib thief at Disneyland 10K - Canadian Running Magazine
2 Likes
Nope. AI.
That’s how they can tag you in photos whilst the event is still going.
Machine algorithm will have spotted two numbers that match in one image, using OCR.
It’ll have then flagged it as an anomaly.
Or, May have passed it to facial recognition first, then realised that for the same number, the face is different. Then flagged it anywho.
#ScienceFacts
3 Likes
Have they found the traditional barrier jumper with a 2hr 30m time to halfway then another 50 mins to complete?
I doubt its that clever. It will be reading the numbers and assigning to the runner, I suspect it double hit on that number and assigned both to the womans account (it seems she was the valid runner and he has nicked it). Cant see they have programmed (at a decent cost) the functionality to detect stolen or duplicate numbers
It seems that London Marathon were alerted via the tweet and not that they worked it out themselves.
2 Likes
There was a guy on the Isle of Dogs who was “virtual runner” or at least that’s what I assumed his number meant
Interestingly 4 members of our running club were running for the same charity and therefore we had sequential numbers and wore the same club top.
At least 2 photos of one of my club mates appears in my photos where his hand/arm obscures the last number on his bib. I guess the AI saw a partial number match and then a kit match and determined that it was me?
1 Like
Or did it assign to all 10 possible people?
1 Like
No way its looking at the kit and pattern matching.
Even in the days of AI, OCR technology is still notoriously flaky so will not always pick the correct number, so rather than leave it orphaned it will assign it to the nearest one it thinks is right.
I was at the police when ANPR came in. The database was full of pictures of road name signs, because they were generally white rectangles with black letters the software assumed they were number plates.
1 Like
I was number 1158 and I’ve got quite a few pictures of people who were 11,58? And the last number has been cut off by the angle they were relative to the camera
1 Like
And this is why we do CaptCha and those image things.
To help train the AI - so it can get better at OCR.
It’d cost next to nothing to knock up what I said.
There’s zero code involved.
It’s all drag and drop.
Easily done in a day or two.
75p per 1,000 photos (I’m assuming there’s less than a million official photos?)
Colour/pattern matching is also a piece of pie.
Zero code again.
It can even match brand logos 
You’ve given me an idea now 
1 Like
You would need some code. Even if you drop all the photos into an S3 bucket. You then need a Lambda function to send them all to Rekogntion. [This is how my DIY ring door bell works] But first you need to send every photo to rekognition to train it so it can recognise when a face matches. If you’re doing it in real time you will need to constantly re evaluate all the earlier photos for matches for all the new photos you’ve just added.
Its not particularly difficult no because AWS (other cloud providers are available) is doing the heavy lifting, but it still needs integrating.
London Marathon are outsourcing the photography. They ask them to provide a system to weed out potential number thiefs with some AI, they say sure. By the time they have added design, dev, QA, PM time London Marathon now have a quote for a good few thousand quid.
Why would they bother, when anyone nicking a number is going to be found out and dobbed in anyway as has happened here.
2 Likes
Totally off topic, but interesting 
Azure has it all built in.
And is drag and drop.
No code.
Face and OCR are both part of the Computer Vision package.
It’s not doing facial recognition, just detection.
Describe what number 11250 is (in JSON)
Compare each JSON description of 11250, quite clearly, they wouldn’t be the same 
TCS (Tata Consultancy Services) partnered with VLM this year - if they’re not doing something like this, they’re missing a trick (or VLM just don’t care about Bib Bandits)
2 Likes
Interesting. Not done much with Azure.
TCS couldn’t get the ballot website working properly to begin with so it might be a bit beyond them. They need us two.
3 Likes
So I’ve just found out that champs times have also changed this year
Now you need sub 2.40 or a 72.30 half!
Well that’s any dreams of a champs start well gone for me now!
3 Likes
Well, with that attitude….
6 Likes
I can’t imagine they had many of them so why cut the times? They’re still fast.
1 Like
That is way beyond my ability! I was a ‘scrape the previous times if it was a good day’ kinda guy. That would need years of proper dedication that I just can’t give. And even then, those times are quick!
Judging by the length of the toilet queues this year the carbon shoe effect has forced them to drop the times to reduce numbers.
*Although with no reference points to compare with I am wildly speculating 
5 Likes
I did actually wonder if it was shoe related
Deffo shoe related