Full AI grading might be a while...
lsica
Posts: 1,747 ✭✭✭✭
SOMEday? Sure. But SOMEday you'll be able to produce rolls of perfect copies of 1804 dollars from your home-replicator. But, for now, for full AI grading, we may need to wait a while. The problem? Just not enough "training data". Sure theres a whole library of True-Views that could be used, but just a glance on this forum shows how many different grades you can get when all you have to judge is a single photograph. What's needed is actual HD VIDEO, but there just isnt enough of that around yet. And also - a lot of the finer intricacies of grading just exists in various peoples heads. They know in their heads whether that extra "eye appeal" on a normally MS-64 coin is enough to make it something that market will say is an MS-65. They know in their heads whether a specific date of a specific variety is almost always flat-struck so you can let more weakly struck pieces grade MS-65 when other dates in that same series require a sharper strike for that grade. They know that for certain dates the planchet handling process at a certain date makes a certain color palate NT, where for another date/mint it's AT. But that information exists between people's ears (or maybe some old obscure Numismatist article) and those that know either can't put it in words - or maybe even don't want to. So, for now, we're going to have to wait for our Graderators where we can put our prizes under a scanner and have it come back and tell us we have an MS64.832 
Comments
The few examples you cite are exactly why we need computer grading. ALL DATES should be graded the SAME. If some dates dont have nice coins surviving so be it. "eye appeal" is subjective so a computer will not include it and it shouldnt . Until we have computers grading coins there will never be any real standards just more of the market grading nonsense like we have now. Markets change. Standards shouldnt
You've assumed a whole lot of questionable facts.
All comments reflect the opinion of the author, even when irrefutably accurate.
The video would have to be very high resolution and be multiple minutes long for it to be viable. Between that, uploading this monster of a file to the AI, and having to analyze it would take much longer than the very short time it takes for a human to grader. AI grading would also remove subjectivity from the equation, which is a very good thing in the world of coin grading. It would also eliminate the need for crossovers or resubmissions, driving down profits and eventually killing grading companies as a whole.
So, it’s never going to happen.
"Shouldn't"? Sure I'll agree with that. But, "in the business we've chosen" (movie buffs
) subjectivity of one sort of another is part of how grading is done. So if you want to discuss what "should" be, we can all sit around and agree. But when we're done yapping and solving the worlds problems and head on in to the coin show together, we're all going to have to deal with that subjectivity. Maybe some day it'll be "fixed". Maybe that'll even happen before we can produce rolls of perfect 1804 dollars with our home replicators. But it's just not how it is right now
Well..... (and IM (useless and) HO)
1) Eye appeal as a criteria in grading is actually in grading descriptions of all the TPGs. No it won't make a MS63 an MS65, and the criteria is just too particular for the uber-grades. But will it make "nicer 63" go 64? Yup.
2) Whole series of coins have special categories (FBL, FH, etc) because they tend to be not super-well struck. And yes even dates. I'm no cross-series-expert, but 1 example that comes to mind are 22-D Lincolns (often not well struck) vs other dates in the series. And if you want the "ultimate" example of (in this case more strict) "grading by date", ask someone who's tried to get an 1880-S or 1881-S Morgan to grade. Pieces that would be MS-whatever "all day" on another date often leave their submitters disappointed.
3) Different NT/AT standards on different coins - send in a toned 1921 Morgan and a toned 1921 Peace with the same color pallet. It's not unlikely that one will straight grade and the other is... well.. "less likely to".......
Strong disagree.
Never is a heck of a long time
... but when AI does finally happen (and again I'm thinking it'll be a loooooooong time) if the TPGs are concerned about future loss of profits from lack of resubs they can say they now have something like "Computer Accurate Grading" and raise prices just enough to make up for any losses. Plus let's be honest about AI, when this does finally happen in the year 2099 or so they'll also save money by not having to pay as many graders.
All of those things can be trained into a machine. Will the machine agree with the human? Probably most of the time. But you would have to say the same for humans agreeing with the other humans.
You also wouldn't need to use existing images to train a machine. It would be relatively easy for a TPGS to create an AI that could grade fairly consistently for submissions. It would be harder to train an AI that could grade outside of the TPGS.
All comments reflect the opinion of the author, even when irrefutably accurate.
The biggest problem, in my mind, is AI can only grade on the photo that has been taken. Garbage in, garbage out. Well, maybe the biggest problem is in the rotation/lighting. In order to grade a coin accurately, you need to see it in various angles, at the proper lighting angle. Take a whizzed coin, for instance. With the right angle and lighting, the coin can look to be a beautiful MS-64. But, rotate it, and you see the whizzing, making it Net EF-40. IF AI sees the first angle, the buyer will pay for MS-64, but could lose on hundreds or thousands of dollars on a damaged coin.
Rocking my "shiny-object-syndrome"!!!
This is not a problem for a TPGS.
All comments reflect the opinion of the author, even when irrefutably accurate.