Quote:
Originally Posted by Snowman
I gave a pretty lengthy overview of the challenges of grading cards using AI in the thread linked below. I work in this field and write similar code to solve other similar problems in a different industry.
The biggest challenge facing AI grading is the ability to accurately grade the surface of a card because typical high-def scans simply aren't adequate for the task (we've all had cards show up in the mail with hidden creases that are not visible from a scan). When TAG arrived on the scene, I was pretty skeptical at first, but I always try to remain open-minded, so I looked into their process more deeply. They utilize an imaging technology known as photometric stereoscopic imaging (they didn't invent the technique, but they did create their own imaging device that uses this) to scan cards with. It allows them to capture the surface from multiple angles using different light sources. Then, they convolve those images into one image using some fancy math. The result is an image that appears 3D and which shows surface flaws like scratches, dimples, tears, creases, etc. Honestly, it's pretty cool, and it's a significant breakthrough for AI grading.
That said, there are still some major hurdles that I'm not sure they'll ever be able to work out for grading vintage cards. It probably works great on modern shiny cards where the variation in surface flaws is much smaller than it is with vintage. But for every card type that they want to grade, they will have to tune an algorithm and taylor it to that card type/set. This is much easier with modern cards that are super cheap and very plentiful. But if they want to start grading low-pop high-end vintage, it's a huge mountain to climb because the AI models they're building will need large training sets of data for each set (thousands of card images if they want a high degree of accuracy) for the algorithms to learn from. Then, they'll have to manually fine tune the results to ensure that the output aligns with market expectations. This is a very labor-intensive process, and it has to be done by a data scientist, and they're not cheap.
I expect to see them slowly dipping their toes into vintage over the next couple years, beginning with sets like 1970 Topps baseball or something like that, but it'll probably be a long time before they ever attempt to grade our more obscure vintage cards. Perhaps they'll take a crack at T206s, maybe 1952 Topps and 1933 Goudeys one day, but I wouldn't expect to see anything out of the mainstream being graded by them any time soon (if ever).
Very cool tech though. Just probably not for what most of us collect.
My lengthier thoughts on AI grading (note, the first post was written before I knew about TAG Grading and photometric stereoscopic imaging, so some of the challenges I mention have been resolved):
https://net54baseball.com/showthread...35#post2132535
https://net54baseball.com/showpost.p...2&postcount=46 |
I working with networking products that use AI. You are 100% correct, you need the training data. Why do you think PSA and others are scanning cards, for their health? The stuff that makes the “cloud” work is all the data centralized in one place. That’s why Amazon has such an advantage over brick and mortar retailers
At this years National, I was at a booth when someone from CSG asked the dealer I was talking to for low grade cards from regional obscure football issues. They wanted them so the graders could get a “feel” for the issue, coloring, stock, size, etc. So, a type collection like Leon’s baseball would have been perfect.
So, in my opinion is Genamint is probably really good at picking up on defects, etc on mainstream issues, but I question what it’s effectiveness would be on difficult issues.
Finally, I spoke to the guy in Toronto for TAG. Only newer cards, but you left out one important thing that the guy mentioned to me. All new cards are usually high grade, however what they are trying to pick up is counterfeits.
Sent from my iPhone using Tapatalk