I don’t think they would use a machine you load the cards into, but lay them on a conveyor belt. Flag the cards that require re-inspection.
It could be surprising low. The food industry use computer graders for fruit and vegetables, now.
Machine assisted visual grading of rare collectibles over the Internet
pdfs.semanticscholar.org/b779/20edbb2c4dad58f3ae525df3ae5651c123f7.pdf
3 Likes
I think people are fixated on the idea of mechanical processing of a pile of cards. The R&D would be exceedingly expensive to do and unnecessary.
All that would be required in terms of hardware would be a decently sophisticated scanner, ideally (but not necessarily) the ability to detect 3d adnormalities such as dents.
Phase 1, scan the cards and have humans grade them, save the scans with the grades. Phase 2, use the labeled data teach a neural network how to grade cards. Phase 3, use the algorithm alongside graders, keep updating the model with new data.
Using a strategy like this, the algorithm in theory should not be picking up on things that wouldn’t be detected by humans otherwise, as a response to Gary’s concern.
One thing that could be appealing about machine graded cards is when they are returned to you it is possible to get a print out of why the card received the grade.
Vertical centering off by 1.5mm
Horizontal centering off by 1.5mm
Left Upper corner 95% of ideal
Right Upper corner 97% of ideal
Left Lower corner 85% of ideal
Right Lower corner 95% of ideal
Color Saturation 75% of ideal
Dents 0
Whiting 0 mm
Scratches 0 - 0 mm
2 Likes
Not really a fan of the machine grading idea but there is definitely room for improvement as far as grading companies go. I would pay more to have a more consistent experience.
Instead of teaching the computer to grade the card, it could be possible to create a perfect virtual card and have the computer identify deviations from perfect. More deviations, lower grade.
I agree that it seems a lot of people are getting hung up on the whole machine handling of cards. I mean for one thing every single card produced was made on industrial printing equipment that is probably a lot rougher than you’d realize with cards. It isn’t human hands gently packing the cards into boosters and boxes. Sure you don’t want a potentially gem mint 5 figure specimen being handled that way again individually on some clunky machine but as PFM said I don’t think it would be needed.
Lay the card on a clean flat surface and get 3D scans. Flip it over and repeat. Initially run it alongside human grading to establish a database and have it flag things the graders may miss. May help in the meantime keeping 10’s with issues slipping through. It could likely tackle the issue that Gary brought up of a machine seeing things humans cannot. The database and learning period would generate the thresholds for corners creases/dings/surface imperfection as obviously no surface is perfectly flat and and looked at on a small enough level every surface would look like sandpaper.
Eventually if it showed enough promise then it could be implemented. Such a large amount of capital required though and it likely won’t happen too soon. I think I have been the only voice lately for PSA to raise prices and I could definitely get behind paying more money for a more precise service. I think they need to raise the prices for current services just to get the demand down for a bit to catch up or increase the supply by training more graders. They’ve had issue with that balance lately clearly. Just dropping off the specials hasn’t been sufficient.
1 Like
I’m glad some of you guys get it.
I also don’t think it would be as expensive as people think and it would pay for itself within a couple of years just like most automated process out there. If grading companies had to charge an extra $2-3 to recoup some of their costs, I’m all for it.
Obviously, grading is very subjective. More so due to the grade being a person’s opinion. While having this 3rd party making the grade certainly helped that issue, the grades are still subjective and a human’s opinion, which unsurprisingly could result in a few mistakes in the grading. Having a uniform system in which the grades are given out, like automated machines doing the grading, could certainly fix that flaw
Believe it or not, it’s probably much easier to “teach” a computer to grade like a human than it would than to define a specific process of identifying a wide variety of different kinds of damage and manually defining how each type of damage should impact the grade. In general, a learning algorithm is far more generalizable and will pick up on nuances a human-defined system may miss
1 Like
Good points.
As far as raising prices to help handle the backlog, that’s sort of what PSA was trying to do by eliminating specials. Unfortunately, as you said, it wasn’t enough.
1 Like
I still object to any system that picks up on nuances a human can’t spot easily then verify with an 8x loop. If it’s success was based on that it would be a complete fail.
You misunderstand what I’m saying here. There are two different ways to develop software that can automate grading.
One is to develop a set of hard-and-fast rules for evaluating a card. Like reducing points for OC or scratching, summing the points at the end and determining a grade. Essentially trying to measure how far the card is from an “ideal” card.
The other method is to use more modern techniques in machine learning to let the software learn what features (that a human grader would pick up on) would cause a deduction in grade.
In the post you quoted, I was arguing in favour if the second method because it is more nuanced and generalizable. For example, take one of those modern hyper rare GX rainbow cards. From my experience, PSA tends to be less strict for centering on these cards (let’s assume this is true for the sake of argument). The first method I described would be very objective and strict about centering and would not treat these cards differently from any other card. However, the machine learning method (if developed appropriately) would pick up on the fact that humans tend to be more lenient on these cards and would treat them likewise.
That’s what I’m saying. Machined-learned software is more nuanced than code written entirely by humans. BUT a 100% machine learning solution would only be able to use features that a human would consciously (or subconsciously) pick up on during grading.
What would it cost from inception to permanent use for each machine? This is really what most of it comes down to. Like I said earlier, many minor surgical robots are being purchased all over the world for 2 Million each plus many additional costs. Well that’s way less than the initial finished product cost to come to market afte r&d etc. That investment made sense though cause thousands will be sold. Not true with a card grading machine.
So again, what would it cost total for that first one to come to market?
In terms of hardware the solution is not very demanding. Maybe put some money to a good scanner, say $2-3k. Probably need a place to store the raw data. $1.5k could get you a 10TB harddrive which should be sufficient. The cost comes in collecting data and developing the algorithm.
You’d have to get someone to scan the front/back of cards. You could pay them $12/hr and it would cost about $25k per year. The data could be used for other purposes such as expediting damage claims since you now have a high-quality scan of the card prior to grading.
After enough data collection, you’d need to hire a software engineer to develop the grading algorithm for you. That’ll run you something like $80-120k per year. I’d imagine to get a decent prototype up and running would take less than a year.
So probably something on the order of $100k. Since it’s a bit of a research project, you could probably apply for government research grants to help mitigate the cost.
Can anybody else here in the business verify this info and the total cost of 100K to get a fully operational “grading machine” up and running?
If so, and it’s accurate, I’d say let’s do it. Make it an E4 business for profit. At such a low cost I’d bet we’d kill it. I’d certainly go a big chunk:)
Interesting replies!
With PSA’s resources, the 50/50 idea seems to be more feasible. As some people said, something like laying cards out on the invisible (glass?) belt with scanners - 3D or rotary. Image processing and feeding into the programs (or AI) to determine a proper grade 
Of course humans will stand by for testing, validating, reporting, etc.
But everything about this feels “off”. This will change the whole imagery of PSA or something like that if you know what I mean. But if they do it, I don’t think they will tell us. They’ll keep it a secret.
Maybe I am looking at this wrong, but if a computer grades like a human, what is its advantage over human?
A well-trained model would be looking for the same thing a human would in evaluating a card. But the probability of overlooking damage is much lower and you’d have more consistency when grading the same card and consistency within a given grade than you would with a bunch of human graders.
1 Like
Humans have taught computers/machines to do many things. Almost all of those things the computers/machines eventually come to do much, much better than us. They could do it with greater speed, accuracy and repeatability.
On a long enough timeline I am sure they will do everything better than us. We can only hope they show us mercy then. * ETA: But for the short term they will likely not show the mercy to some cards that flawed human graders currently do. Those “weak 10’s” or worse completely misgraded 10’s everyone hates.
2 Likes