Forums Latest Members

Using A.I. and Microscopic Photography to Authenticate Watches - Coming Sooner Than You Think!

  1. WatchVaultNYC Aug 19, 2018

    Posts
    3,719
    Likes
    4,189
    A few weeks ago, I was in Manhattan for a business meeting with a company with whom I had partnered with for the pre-owned luxury handbag business I started up with a partner a few months ago. The company was basically an online authentication service. By some estimates, 9 out of 10 luxury handbags in the market are fake, so it's imperative that when you're in the business of selling pre-owned handbags worth thousands of $$$, you have the highest level of certainty that what you are picking up, and what you are selling, is real.

    They give you this device (which we now use) that takes dozens of microscopic photos of a handbag, which is then sent to a server to compare with hundreds of thousands of photos of legit handbags. Using A.I. (artificial intelligence) they decide "yay" or "nay", or "not sure". Having a background in computer science, and AI in particular, this is a sound concept and widely in use now in ways that have become invisible to us - Google search and auto-complete, Alexa, Siri, Facebook ads, Tesla autopilot, etc, etc.

    Right now, this company could be considered the premiere online authenticator for luxury handbags. But here is where it gets interesting.

    During the meeting, they showed me their skunk works office, where I saw piles and piles of vintage sneakers.. and watches. They were going to attempt to do the same thing for watches! And I was very excited - I was not there for watches, but now I was. Clearly the AI/micro photography concept as it is currently implemented and a just a bit of tweaking could identify fake parts, redials, flagged serial numbers, and frankens. I mentioned to them the issue of severely polished watches which could cause mis-identification, and also the real issue of very rare watches - how do you get enough photographic samples for an AI to make a sound decision? They were of course still working that out.

    I cannot divulge more than this at the moment, but will let you know more as soon as I am allowed to. It's very exciting stuff!
     
  2. WhereMadnessLies Aug 19, 2018

    Posts
    179
    Likes
    983
    Very cool. It would be fascinating to know the methodologies behind this sort of system. I did a degree with computer science and AI but concentrated much more on optimal route finding.

    I suspect once they get the tech up and working it could be expanded to many different collecting markets....baseball cards, coins etc...
     
    WatchVaultNYC likes this.
  3. Dan S Aug 19, 2018

    Posts
    18,604
    Likes
    42,878
    I suspect it is a neural net-based approach, in which case you will probably not get any information about why a given handbag classified as authentic or not authentic, it will just be classified that way by the algorithm. My collaborators and I use some similar approaches in our research, and with good training sets, the approach can be pretty accurate, but it doesn't give much insight.

    Of course, this is just a guess, and they might be using an entirely different machine-learning approach. I'm not sure I would really call this AI, but that's a trendy term.
     
    WatchVaultNYC likes this.
  4. w154 Aug 19, 2018

    Posts
    2,532
    Likes
    5,454
    I work with optical measurent guys and certainly you can do interesting stuff if you have good equipment. The obvious thing would be to measure the degree of material lost on a polished case versus a NOS case. You can measure everything up in a few minutes and create a 3D model for each one, then overlay them and see the material lost in all locations. You just need a reference surface where there’s no wear which could be the inside back of the lugs (or ideally an interior surface if the cases are stripped down). You can measure to around 3 microns precision with ours.
     
    WatchVaultNYC likes this.
  5. kilofinal Aug 19, 2018

    Posts
    361
    Likes
    368
    wow......you mean we won't see any more - " can you tell me if this watch is fake?" posts?
     
    WatchVaultNYC likes this.
  6. MikeMan2727 Aug 19, 2018

    Posts
    1,653
    Likes
    8,682
    Very cool! Although it might be more difficult with vintage watches since there aren't many available examples of certain references to use as a benchmark.

    I can see this being very helpful with modern watches, however.
     
  7. w154 Aug 19, 2018

    Posts
    2,532
    Likes
    5,454
    Edit just to give an example. This is a overlay of two models. The reference model is probably a drawing, and overlayed is a real part. So you can see how it deviates. As there are no drawings available for old cases you’d have to use a NOS case as a reference. Anyway, all over the surface you can see by color how much the measured part deviates from what it should be. Very cool, and perhaps worthwhile for validating very expensive watch cases.

    9A838871-2B3A-4813-850D-DB8A509D2559.jpeg
     
    Edited Aug 19, 2018
  8. Ninja2789 Aug 19, 2018

    Posts
    488
    Likes
    691
    Man. This an incredibly interesting concept. So the first question I would ask is how do we get a decent enough sample size of correct vintage watches? As I understand it, AI recognition works best when you can identify a base case (eg a correct handbag) and compare multiple examples agains it.

    For example, I would imagine the handbag market is filled with examples of bags that are consistent in nature. You can find multiple copies of a correct bag and basically compare it against the correct example.

    However, the vintage market has an incredible wide range of conditions that make me think it'll be all but impossible to compare the different conditions of various vintage watches. Tropical dial vs damaged dial vs complete fake dial? There will have to be an assumption made and that may damper the accuracy of the watch.

    So what I would say is to invert the AI. Don't look for correct examples. Use known FAKE examples and then provide a % of confidence that you know it's fake. I think that would make it much easier and consistent data to compare.
     
    WatchVaultNYC likes this.
  9. drhombus24 Aug 19, 2018

    Posts
    531
    Likes
    836
    The other day I was playing around with a few machine learning libraries in python, and it pretty cool to see what’s possible with a lot of good data. It’s a very exciting time to be a student!
     
    WatchVaultNYC likes this.
  10. Foo2rama Keeps his worms in a ball instead of a can. Aug 19, 2018

    Posts
    17,045
    Likes
    25,211
    I hope they get good results. But my experience in what an AI can do and not do... this is a tough one. Perhaps for specific models maybe like Subs
    ... but the huge amount in variation in 60’s Seamasters it seems hard.
     
    WatchVaultNYC likes this.
  11. Vercingetorix Spam Risk Aug 19, 2018

    Posts
    3,238
    Likes
    5,219
    Also the variation in case/dial manufacturers. Omega had what a dozen Swiss and half a dozen US case manufactures in the 60s?
     
    WatchVaultNYC and Foo2rama like this.
  12. auxpomme Aug 19, 2018

    Posts
    269
    Likes
    248
    Sign me up!
     
    WatchVaultNYC likes this.
  13. shishy www.hpmor.com Aug 20, 2018

    Posts
    1,006
    Likes
    2,894
    I had a similar idea but getting a good training set was really difficult. Kudos to them if they managed to pull it off in a reliable way.

    Will keep an eye out for this.
     
    WatchVaultNYC likes this.