Hi! Having been the beneficiary of this community's many active and gracious contributors over the past 2 years, and having been so new I couldn't contribute anything back, I thought I might offer a little something. I have been using CoinSnap (about $30 yearly) for about 3 years. Three years ago, it was an interesting app that more or less could "get a new guy in the ballpark" many times, but it was way way off sometimes. I've noticed it has improved greatly over the past 3 years, and I wanted to give it an acid test. I had some coins graded by ANACS recently and wanted to share the results of ANACS grading vs. the CoinSnap grading. Understood that ANACS isn't PCGS or NGC, but as most of mine were copper with some varieties, ANACS seemed to be the most cost effective fit. As you can see from the table below, ANACS and CoinSnap agreed 7 out of 13 times. They were close 5 of 13 times and they disagreed strongly once. In 3 of the 5 times where they were close, the grade was within the tier or right next to it, so I called that "acceptable." In 2 of the 5 times where they were close, it was a "two-step jump across adjacent tiers," which I thought would be acceptable, but you may not think so, and that's OK. For a person still learning though, I'll take it. The time they disagreed was for a 1785 VERMONTS copper that was holed and bent, so on that one, I'll give CoinSnap a little slack as it was more conservative than ANACS. I was surprised it could identify it at all as it took me a couple of hours initially to figure it out (I wasn't orienting it correctly). CoinSnap won't tell you that the grade is "details." However, it will tell you that it's scratched or corroded and where that is located. It will tell you about the luster or lack thereof and also about the wear, if any, and where that wear is located. So, you can pretty much figure out if it is likely to generate a "details" grade or not. It will also pull recent sales of that coin in that condition and will give you a value. You can also add that coin, with pictures, grade and value to your CoinSnap online collection and pull it up anytime on your phone. Up until about a year ago, different lighting on the coin would generally generate a different grade. That seems to have been resolved, although I can still get it to do that once in a while if I really mess with the lighting to try and trick it. In case you are thinking that this post is an ad for CoinSnap, I assure you it's not. I wanted to share with this community my experience with it as I've scanned a couple thousand coins with it so far. However, this is the only time it dawned on me to really test it out against a TPG and to try and give back a little something to CT. These coins are old and most have "details." As I work with CoinSnap on more modern coins - Morgans, Mercury Dimes, Buffalo Nickels, etc., I double check the results with the PCGS website as well as CG and the results aren't better or worse that what I've shared - it keeps me from being too optimistic! So, here are thirteen data points in the chart below for your consideration (or not). Thanks!
Interesting. I think it starts to run into more trouble with AU/MS coins. I wonder how "CoinKnow" would compare.
Have you tried it with some of the guess the grades on this site? See how the app does on some higher grade coins? Pictures might not be the best, but a good way for you to practice with this app. I have not tried that with chatgpt - so far I have only tried it with my coins. it is not exact and hits a few of them. Variety attribution can be a challenge for it, but I can usually guide it to the correct attribution - but I have to know it first. I also wonder how these AI things learn or if they only do that when the code changes. We all know EBAY AI is very limited to what it does and who it favors. I doubt they will ever be exact especially since two coins can look vastly different in pictures - even in really good pictures - and still have the same grade.
A couple months ago my brother was playing around with a coin app. Not sure if it was coinsnap or not. I got out some PCGS and NGC graded coins for him to try it on. And that was definitely the case, with the AU/MS coins. By the way, thank you @Atlanta for sharing your observations.
From what you say it seems to be improving. I would not consider G4 coins getting VG8 “acceptable “ though. That’s a big jump in value for a lot of coins.
Thanks for posting but it appears the tech/algorithm are in an early stage. If one is looking for reasonably right appraisals, this might be of some use. I don’t think that it’s the product for me though.
With coin grading always being contentious, even with the top TPGs, I wouldn't expect an app to do any better. The aspect that would interest me the most is if it could accurately detect counterfeit coins, that would make it truly worth using IMO.
Hi - I wanted to follow-up on some of the questions raised and my next steps... @physics-fan3.14 - CoinSnap is an AI app that can identify and grade coins @Alabama Stacker @AdamL - Early on, CoinSnap would just say "AU" or "MS" which, of course, isn't much help. Within the past 12 or so months, it is now giving me a numerical grade such as AU-58, MS-61, MS-62, MS-63, etc. I had a $5 come back from PCGS as PR69DCAM - I had tried to grade it using CoinSnap through the mint's plastic capsule, but it didn't work well and I didn't want to break it out of the capsule myself fearing I would screw it up. I have a number of higher grade coins (AU/MS) now at the TPG and graded them via CoinSnap before I sent them. I will report back to the community how those results stack up against the TPG grade. @mark_h - I have tried it with pictures on the internet on known graded coins, but it is consistently inconsistent. It will identify the coin, but sometimes the CoinSnap grade agrees with the TPG, sometimes it's a bit too high and sometimes it's a bit too low. It seems to need to actual coin to give it the best shot at correctly getting the grade. @ksmooter61 - for the one counterfeit that I have (or at least know that I have) it can't tell if it's real or not. It scanned it and gave a grade without mentioning it was counterfeit. In a couple of months, I'll post another thread about the higher graded coins from the TPG and how CoinSnap graded them before I sent them off. Stay tuned for the TPG Vs. CoinSnap SMACKDOWN!
To validate software like this to a significant degree you would need a large number of replicates…likely hundreds…and various challenges based on the scenarios you develop. I’m not sure of your final aim but as long as you’re having fun, go for it.
@Kevin Mader - yes, agreed. I'd need a statistical sample to get to a +/-5% confidence and that would be hundreds of coins. But, an interesting, fun exercise that I thought I would share.
It’s a fun experiment from my perspective. I think the community here would be interested in the outcome. Just putting it into the minds of our fellow enthusiasts that there are limitations; essentially curbing some expectations as we all might have a different application for the results. From the comments here I think that it’s understood. Have fun!
I look forward to future results. I have noticed with chatgpt it typically does not call a coin a fake , but seems to say I think it is a fake, but should be confirmed with a tpg. I think one of the issues AI will have is weak strike vs wear with just pictures. Hard enough for me at times.
