Card grading and identification Ai

[maniac_73]

Card not cars lol can’t edit title typo

. Whipped this up quickly in ChatGPT for fun. Try it out and tell me what you think https://chatgpt.com/g/g-6Xukfe94M-grademycard


Sent from my iPhone using Tapatalk

[Leon]

Originally Posted by maniac_73

Card not cars lol can’t edit title typo

. Whipped this up quickly in ChatGPT for fun. Try it out and tell me what you think https://chatgpt.com/g/g-6Xukfe94M-grademycard


Sent from my iPhone using Tapatalk

Title fixed.

To edit the title of threads ..go to edit it and look on the lower right for the words "go advanced"...click on those words and you can edit the title.

Also, it looks like we have to sign up to read it?

[rdwyer]

I like it. Uploaded a front and back and it gave me all kinds of info on the card.

[G1911]

I have tried to use GPT 4o for cards before, and it has mostly failed and returned false information for me, when using simply the basic chat without any preprogramming or memory saved knowledge. I am fairly impressed this time with this tool here. I did 3 tests in my boxing realm, before hitting the image limit on my free personal account. I tested with 3 cards pretty much randomly pulled from my boxing archive, making sure each is a decent quality picture and that it is from a known set documented on the internet, so that GPT could reasonably get it right if it was able to reverse image search and connect to similar results online, and then use that to pull sales data and identifications. Obviously no AI tool is going to be able to pull data for undocumented cards.


Test 1 - wrong. I uploaded an E80 caramel card, which it claimed was a T218 (T218 looks very different and doesn't have multiple subjects on a card). GPT does this a lot - if it isn't sure but returns some results, it just picks a claim and runs with it instead of admitting it doesn't know. It did correctly read the names and then pull information about them both, which was cool. The pricing information was, of course, just completely made up and meaningless.

Test 2 - Passed! Correctly identified it. "Shifted slightly to the left" and the grade on this miscut weren't very accurate, but it did identify the card.

Test 3 - This was supposed to be the hard one of the three designed to fail, but possible to succeed. N332 is a known set, but Mitchell here only surfaced with a full recognizable copy of the card with the ad a few months ago, not giving it much to pull from the internet. I am quite impressed it got it right. It also correctly defined it as "rare" in the verbiage, where it just praised me for the other 2 cards that are not rare. It gave the same pricing information for N310 and N332, which looks like GPT just making things up like it often does when it can't quite pull enough data but doesn't find 0 data.


So it identified 2 of the 3, 2 of the 3 had accurate grading and for the failed grade (N310) it still identified many of the condition issues, and the pricing didn't seem to really work. Pretty good, considering that I am giving it recognizable but not famous cards that most collectors wouldn't use. Whatever it can do for N332 is likely to be better for a 1956 Topps. This is cool

[maniac_73]

Thanks Guys, I really am impressed with the quality of results versus some of the older models. As I keep fine tuning it I think it will even get better. Definitely will never be perfect but a fun little project for me as continue to explore AI.

[Goudey]

did you fine-tune 4o is that the base model with a knowledge base?

[G1911]

Having fun with this, I wanted to throw it another possible but difficult curveball. Based on my test yesterday I have no doubt it will recognize a 1954 Bowman Willie Mays I show it or something 'regular'.

I uploaded an uncut sheet (actually a series of panels put together to recreate the sheet) of T220 Silvers, and gave it no context at first. While wrong, I am again impressed. It correctly identified it is looking at a sheet, it jumped at the tape measure clue to piece together more information, and it correctly identified that these are T cards. It also knew that any tobacco sheet is a very rare item. While this piece is authentic, by recognizing it is unique or near unique the tool reported back with authenticity check information to validate, instead of assuming it is real, seeming to recognize its own limits. The grading instructions were also a little different, seeming to understand that these panels cannot just be submitted for encapsulation. It wrongly assigned T206, which I have noticed GPT does a lot with old cards - if it can't figure out a cigarette card it tends to just run with T206.


I then tried to help it, calling out that the cards picture boxers and have silver borders, so it's not T206. GPT then went completely off track and posted a ton of false information it made up about the 1951 Topps Ringside, just picking a boxing set at random. Hallucination. This is where I stop understanding the models at all, as there is nowhere it could search online to come up with these things - the claims are things it has invented.


The results I am getting are markedly better than when I just ask GPT the question without whatever pre-programmed information or context the OP gave this model he made. While not quite right, it is a difficult 'trick' and it understood a lot of my trick, and reshaped its format of response around the item.

[maniac_73]

Originally Posted by Goudey

did you fine-tune 4o is that the base model with a knowledge base?

Fine tuned the 4o with prompting. Its a huge leap over base. As you can see with G1911's experiments it still does suffer from the odd hallucination but its getting good

[Leon]

Originally Posted by maniac_73

Fine tuned the 4o with prompting. Its a huge leap over base. As you can see with G1911's experiments it still does suffer from the odd hallucination but its getting good

"Good" is relative. If perfected, is that a good thing?
.

[sports-cards-forever]

This is fascinating. I think it's the future

[Snowman]

I uploaded a few cards. It's not good at judging the centering for some reason.

[maniac_73]

Originally Posted by Snowman

I uploaded a few cards. It's not good at judging the centering for some reason.

You're right. Just added PSA centering standards for it to use as a guideline. Seeing if that fixes it

[Bigdaddy]

I uploaded a 1964 Topps Sandy Koufax and it identified it as a 1966 Topps. Most of the information seemed to be in the ballpark except for the year of issue.

[prestigecollectibles]

I tried doing a Japanese baseball card
1949 JF 35i Asayama Fusen Gum Type 8--Kaoru Betto

I was pleasantly surprised by the results. It shows great potential. Although it couldn’t recognize the player or team names written in Japanese and the card isn’t a menko, I’m excited to see how it develops in the future.

[Snowman]

Originally Posted by maniac_73

You're right. Just added PSA centering standards for it to use as a guideline. Seeing if that fixes it

I'm not sure how it works (I haven't looked into it), but if it's a LLM algorithm, then it might just be flagging every vintage card as being off-center because 95% of vintage cards are in fact off-center and it can achieve 95% accuracy by simply guessing that. It does not appear to be actually measuring the borders. Perhaps that can be fixed though.

[steve B]

I would think centering would be easy. But probably also has to be specifically programmed.

[Snowman]

Originally Posted by steve B

I would think centering would be easy. But probably also has to be specifically programmed.

It's pretty easy with convolutional neural networks (CNNs). You use them to detect where the borders end and begin (called edge detection) on each side and then count the pixels between the edges. But it's clear from the results that it's not doing that, so that's why I suggested that it's doing the typical ChatGPT LLM approach and simply bluffing because most cards are in fact OC.

[maniac_73]

Originally Posted by Snowman

It's pretty easy with convolutional neural networks (CNNs). You use them to detect where the borders end and begin (called edge detection) on each side and then count the pixels between the edges. But it's clear from the results that it's not doing that, so that's why I suggested that it's doing the typical ChatGPT LLM approach and simply bluffing because most cards are in fact OC.

Its using the OpenAI LLM. Not using Image Recognition transformer. The parameters are just a percentage of each side. Would be interesting to see how an Image Recogntion LLM would work as you're right it does the recogntion by pixel which would in theory be a much more accurate method when fine tuned properly.
This is a just a Custom GPT I was playing with and am tuning as I go. A quick thing for fun to see what is possible.

[Leon]

Originally Posted by Bigdaddy

I uploaded a 1964 Topps Sandy Koufax and it identified it as a 1966 Topps. Most of the information seemed to be in the ballpark except for the year of issue.

It's only a matter of time for a lot more refinement in the AI area. It's in it's infancy still. I have mixed feelings on it, myself.
.