Posted a couple Saints as a GTG earlier this week - they have been added to the OP in this thread.
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Have been tracking the "serial" numbers on the reverse from coins I've seen or own. Here's an example of a higher 1st number with variable following numbers. 1st number unlikely to be a sequential invoice number - could it be a customer ID number??
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
I bought my Compugrade slab including Compugrade sample slabs over 10 years ago. I currently have a dozen of them. The sample Compugrade were half dollars and they set me back $10 each.
All of the non sample ones are Morgan silver dollars as I recall.
Those Saints are homeruns in the Comnpugrade slabs!
@Lakesammman said:
Would anyone mind if I changed the title to make this a Compugrade "master" thread where Compugrade info can be accumulated "going forward" (man, I hate that admin jargon..... )
I personally do not think it matters as long as the word COMPUGRADE is in all caps.
@ricko said:
I have not seen any other coins slabbed by Compugrade.. and those only pictured here...It was an effort made before it's time.... now that AI is being perfected, I think we will see computer grading and real standards implemented... perhaps in the next 5 to 10 years. It will have a major impact on the hobby and the grading business....However, in the long run, it will benefit the hobby. Cheers, RickO
I disagree with it helping the hobby. We just have "collectors" that can read numbers on a label, but not able to grade coins. The grading of coins with become a lost art and black magic.
@ricko said:
I have not seen any other coins slabbed by Compugrade.. and those only pictured here...It was an effort made before it's time.... now that AI is being perfected, I think we will see computer grading and real standards implemented... perhaps in the next 5 to 10 years. It will have a major impact on the hobby and the grading business....However, in the long run, it will benefit the hobby. Cheers, RickO
I still think it's very unlikely.
The kind of AI that is making strides today is machine learning image classification, where we use a set of tagged images to self-program a neural network. This isn't hard, you can do it at home on a Raspberry Pi to differentiate between types of garbage (recyclable or not).
Those models are dependent on having a training set of properly labeled images. The number of images used to train the model impacts the ability of the model to identify a new image.
It might take 10 or 20 images to differentiate between a recyclable can and some food waste. It takes 100s of images of dogs in all orientations, all sizes, all breeds, different image quality, and 100s of cat photos to train the model to differentiate dogs from cats. Interestingly, we can't say how the model works to differentiate the two. But it works.
In the real world, these systems are highly impacted by small changes in the image - there are a number of papers on how tiny modifications of an image (indetectable to a human) lead to misidentification.
Could we program a model to tell us what type of coin something is - certainly. It might take a lot of tagged images if we go back to the era of hand-cut dies, but a basic set of US circulating coins wouldn't be hard. The 50SQ and ATB and Innovation dollar programs up the number of images required, but they aren't hard to get ahold of.
Thus in order for a ML system to grade coins, we need two impossible things:
A large set of images of coins where everyone agrees on the grades so they can be properly tagged.
Nobody altering images to "game" the system
Yeah right.
Now all is not hopeless. Saying you can't build a general ML model to grade coins from random quality photos is not the same as saying you couldn't build a system to grade specific coins like our hosts.
They already have a set of training photos taken under similar conditions, namely their photo archive. By controlling the imaging process, they eliminate adversarial modifications and many of the issues regarding image quality.
For many coins in many series, they have enough tagged images to train an ML model. Not a system that can grade any arbitrary coin, but one that could grade say an untoned Morgan. This isn't even difficult.
-----Burton ANA 50 year/Life Member (now "Emeritus")
@Fisher25 said:
It took me a while to finally be able to post on this thread, but I have something that may be interesting to some people on here. What do you think?
@Lakesammman said:
Anyone have a Compugrade Saint or picture of one to share??
I think it's a Saint in a Compugrade slab. Tim is going to be happy!
@Fisher25 said:
It took me a while to finally be able to post on this thread, but I have something that may be interesting to some people on here. What do you think?
@Lakesammman said:
Anyone have a Compugrade Saint or picture of one to share??
I think it's a Saint in a Compugrade slab. Tim is going to be happy!
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Any new Compugrade purchases?? I'm seeing a few now and then on ebay, but haven't found a grade I need for the set in quite a while!
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Yep, they seem to show up really occasionally, but the prices are such that I'm not really chasing them. It's not a primary focus and we never really figured out the labels (which is really what I care about).
-----Burton ANA 50 year/Life Member (now "Emeritus")
John Kraljevich offered a Compugrade brochure on Facebook last night. Due to my ignorance on how auctions work there, I was outbid. John gave me permission to post these pictures.
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Abstract
A method and system for accurately and objectively evaluating the numismatic quality of a test coin and/or for fingerprinting the test coin for purposes of identification is disclosed. Central to both the grading and fingerprinting aspects of the invention is the exact, numerical evaluation of any detracting marks on each side of the coin. In particular, each detracting mark on the coin is identified, located and measured. An assigned quantity representative of the detracting significance of each mark is then calculated by adjusting the measured surface area of the mark by a factor representative of the relative grading importance of the area on the coin where the mark is located. The assigned quantities and corresponding mark location identifiers are stored as a unique test coin fingerprint. The grading aspect further requires that the assigned quantities for each side be separately summed and correlated into a grade via comparison with a preexisting database of values representative of numismatic grades. Also, the method and system preferably include an automatic analysis of each coin side surface to determine a mint luster value, surface wear value, strength of strike indication, and whether artificial treatment of the coin has occurred.
Image classification as mentioned above would be the answer for identifying what type of coin. However the way classification works would not be very good at grading. It would be identifying features and only those within the training set.
A novel approach that would have promise would be a Siamese neural network. If you train it on images representing wear states I’d say it would do a pretty good job at seeing how similar or not a new object is. And it would have the advantage of when seeing new items it could then be held off and the embedding later labeled and then become part of the comparison set.
@TurtleCat said:
Image classification as mentioned above would be the answer for identifying what type of coin. However the way classification works would not be very good at grading. It would be identifying features and only those within the training set.
A novel approach that would have promise would be a Siamese neural network. If you train it on images representing wear states I’d say it would do a pretty good job at seeing how similar or not a new object is. And it would have the advantage of when seeing new items it could then be held off and the embedding later labeled and then become part of the comparison set.
Actually, as patented - 36 years ago (1987 application) - it was for grading. You are almost looking backwards in time and saying the Wright Brothers should have used GPS to document their first flight.
in the mid 1980s, Neutral networks were a research topic that has basically run its course with the computers of the day and was just being revived as smaller computers became available... When I took CS231 Topics in AI from Jim Low in 1980, NN was just a historical oddity. He did however bring popcorn to the Friday afternoon movie screenings which is why most of us took the course.
[I did a project for Jim implementing (name,type,value) tuple functions in SAIL]
It's not until the 90s that image processing enters the field and at the time, only with what is today considered toy images (handwriting samples 32x32 pixels of gray scale).
What is the first paper on Siamese Neural Networks? 1994??? It might be a far better choice than today's throw spaghetti (as many images as you can grab) against the wall and see what sticks... but it's not the sexy stuff that gets the Unicorn valuation for recreating Eliza with more data...
-----Burton ANA 50 year/Life Member (now "Emeritus")
@joebb21 said:
Man- those Compugrade ms63 saints look like ms65 in todays time.
proof inflation exists here too!
Hard to say. It was never proven, but widely said that the grading didn't work and the few coins they did slab were human graded. Perhaps with some input from the automation, but...
Certainly the patents are very technical grading ... count the # of dings, weight the counts based on a map of how important the area is to the overall coin, compute a #. And we all know how well technical grading has faired against today's market grading schemes.
The difficulty in creating the maps might indicate why only a very few types of coins have ever been seen in CompuGrade slabs. It was never going to be a general-purpose grading mechanism.
-----Burton ANA 50 year/Life Member (now "Emeritus")
I'm curious about how that system they built would grade today if you only changed the image acquisition device and added processing power to handle more data. In 1991, they had to use a video camera with and A/D converter to generate what I imagine was a 512x512 image of the coin. That image chain probably had problems with signal overshoot where there were points of very high contrast. Not sure how well they'd have been able to be filtered out. Pull that out and put in a modern DSLR that can produce an image of at least 2048x2048 (16x the information) and see what happens to the grading results. Do they stay the same? Do they consistently shift in one direction? Do they fluctuate a lot? Do they converge more on "the standard"?
@TurtleCat said:
Image classification as mentioned above would be the answer for identifying what type of coin. However the way classification works would not be very good at grading. It would be identifying features and only those within the training set.
A novel approach that would have promise would be a Siamese neural network. If you train it on images representing wear states I’d say it would do a pretty good job at seeing how similar or not a new object is. And it would have the advantage of when seeing new items it could then be held off and the embedding later labeled and then become part of the comparison set.
Actually, as patented - 36 years ago (1987 application) - it was for grading. You are almost looking backwards in time and saying the Wright Brothers should have used GPS to document their first flight.
On the contrary. My point was not at all about what could have been done or anything like that. It was about what might be a solution to doing it today.
@messydesk said:
I'm curious about how that system they built would grade today if you only changed the image acquisition device and added processing power to handle more data. In 1991, they had to use a video camera with and A/D converter to generate what I imagine was a 512x512 image of the coin. That image chain probably had problems with signal overshoot where there were points of very high contrast. Not sure how well they'd have been able to be filtered out. Pull that out and put in a modern DSLR that can produce an image of at least 2048x2048 (16x the information) and see what happens to the grading results. Do they stay the same? Do they consistently shift in one direction? Do they fluctuate a lot? Do they converge more on "the standard"?
I think that's a big issue.
The Compugrade patent references an Image Analysis System that could do 500x magnification as an example of the imaging system that might be used, and says
It will be apparent to those skilled in the art from the following discussion that other types of imaging hardware and/or systems may be utilized in implementing the present invention. For example, scanning electron microscopes, energy dispersive spectrophotometers, VCRs, laser scanners, holography, interferometry and image subtraction are a few of the alternate, presently available types of equipment technologies which may be used.
Interestingly, the 2nd & 3rd Compugrade patents were about lighting the images to reveal the surface of the coin...
Even the best imaging system today is nowhere close to 500x lifesize for a full coin. Kodak is quoted as a full frame 35mm image is 6K resolution or something like a 20MP (5568 × 3712 pixels) image. The largest picture ever taken is usually quoted as a NASA 1.5 billion pixel image (69,536 x 22,230).
BTW, a modern mirrorless shoots raw at 30, 45 (8192 x 5464), even 60MP... Yet when DIY images first came to the web, and some people had - gasp - 640x480 cameras, the 'standard' was 320x240. People actually cropped to that, discarding the original image, "because we'll never need higher resolution".
Today we know better... we know imaging books & things at 4k or higher resolution won't suffice forever.
But what does happen when you have to reimage your entire training set? Can you even find the same coins? Must you own them, store them, and plan to re-image them every 5 or 10 years???
What happens when that image isn't raw but is processed through an AI designed to make better people & landscape shots??? Does the camera wait for Miss Liberty to look straight at the lens and not blink? Does it glam up her photo?
What happens when the image sensors change and the color response to even the same lights is different? One of the improvements being worked on is to stack the sensors so you don't have little triangles of RGB sensors but actually read the values at the exact same spot.
-----Burton ANA 50 year/Life Member (now "Emeritus")
@TurtleCat said:
On the contrary. My point was not at all about what could have been done or anything like that. It was about what might be a solution to doing it today.
Ah. Run with that... what do we do when the answer isn't to get more labeled images to stuff into the training set? Because the more images, the more people doing the labeling the more disagreements you will have.
And what's true for one series, even for one date range of a series isn't true for others?
We had a preso on Buffalo Nickels two weeks ago at coin club and the presenter was talking about the difficulty in grading the late 20s early 30s coins. We've all seen those gorgeous 1936 and 1937 coins when new equipment was finally able to bring out the details from the dies. But what about say that 1927 coin that never had the details well struck up and you might mistake a MS coin as having the same "look" as a VF 1937.
Do you apply the same standards to 1880/1881 S Morgans (when San Fran was just dialed in and turning out gorgeous coins) to the late New Orleans coins?
-----Burton ANA 50 year/Life Member (now "Emeritus")
@TurtleCat said:
On the contrary. My point was not at all about what could have been done or anything like that. It was about what might be a solution to doing it today.
Ah. Run with that... what do we do when the answer isn't to get more labeled images to stuff into the training set? Because the more images, the more people doing the labeling the more disagreements you will have.
And what's true for one series, even for one date range of a series isn't true for others?
We had a preso on Buffalo Nickels two weeks ago at coin club and the presenter was talking about the difficulty in grading the late 20s early 30s coins. We've all seen those gorgeous 1936 and 1937 coins when new equipment was finally able to bring out the details from the dies. But what about say that 1927 coin that never had the details well struck up and you might mistake a MS coin as having the same "look" as a VF 1937.
Do you apply the same standards to 1880/1881 S Morgans (when San Fran was just dialed in and turning out gorgeous coins) to the late New Orleans coins?
This specific question came up at their presentation 32 years ago, comparing 81-S to 92-O. Their response at the time was that the rules had to be developed for each date and mint individually to take these things into account. Mind you, no AI then like there is today. Simply weights on the degree of deficiencies of the surfaces, luster, and strike.
@messydesk said:
I'm curious about how that system they built would grade today if you only changed the image acquisition device and added processing power to handle more data. In 1991, they had to use a video camera with and A/D converter to generate what I imagine was a 512x512 image of the coin. That image chain probably had problems with signal overshoot where there were points of very high contrast. Not sure how well they'd have been able to be filtered out. Pull that out and put in a modern DSLR that can produce an image of at least 2048x2048 (16x the information) and see what happens to the grading results. Do they stay the same? Do they consistently shift in one direction? Do they fluctuate a lot? Do they converge more on "the standard"?
I think that's a big issue.
The Compugrade patent references an Image Analysis System that could do 500x magnification as an example of the imaging system that might be used, and says
It will be apparent to those skilled in the art from the following discussion that other types of imaging hardware and/or systems may be utilized in implementing the present invention. For example, scanning electron microscopes, energy dispersive spectrophotometers, VCRs, laser scanners, holography, interferometry and image subtraction are a few of the alternate, presently available types of equipment technologies which may be used.
Interestingly, the 2nd & 3rd Compugrade patents were about lighting the images to reveal the surface of the coin...
They emphasized the importance of having multiple views of a coin during the presentation, with each view providing specific input that would be used for grading.
The "500x" magnification number is somewhere between silly and meaningless. You see stuff like that on cheap digital microscopes today. Anything can be blown up to an arbitrary magnification. The image might (will) be crap, but at least you have a big number in front of the 'x'.
But what does happen when you have to reimage your entire training set? Can you even find the same coins? Must you own them, store them, and plan to re-image them every 5 or 10 years???
The bigger issue is that what do you do when your grading scale is rejected by the marketplace? That's a potential fatal bug in your training, so you have to start over with a new training dataset that is perceived as being the most reflective of the current market. Then you end up with grades that change with versions of your software, killing the repeatability claim you touted. The alternative is that you stand fast and nobody uses your service anymore.
@ricko said:
I have not seen any other coins slabbed by Compugrade.. and those only pictured here...It was an effort made before it's time.... now that AI is being perfected, I think we will see computer grading and real standards implemented... perhaps in the next 5 to 10 years. It will have a major impact on the hobby and the grading business....However, in the long run, it will benefit the hobby. Cheers, RickO
I still think it's very unlikely.
The kind of AI that is making strides today is machine learning image classification, where we use a set of tagged images to self-program a neural network. This isn't hard, you can do it at home on a Raspberry Pi to differentiate between types of garbage (recyclable or not).
Those models are dependent on having a training set of properly labeled images. The number of images used to train the model impacts the ability of the model to identify a new image.
It might take 10 or 20 images to differentiate between a recyclable can and some food waste. It takes 100s of images of dogs in all orientations, all sizes, all breeds, different image quality, and 100s of cat photos to train the model to differentiate dogs from cats. Interestingly, we can't say how the model works to differentiate the two. But it works.
In the real world, these systems are highly impacted by small changes in the image - there are a number of papers on how tiny modifications of an image (indetectable to a human) lead to misidentification.
Could we program a model to tell us what type of coin something is - certainly. It might take a lot of tagged images if we go back to the era of hand-cut dies, but a basic set of US circulating coins wouldn't be hard. The 50SQ and ATB and Innovation dollar programs up the number of images required, but they aren't hard to get ahold of.
Thus in order for a ML system to grade coins, we need two impossible things:
A large set of images of coins where everyone agrees on the grades so they can be properly tagged.
Nobody altering images to "game" the system
Yeah right.
Now all is not hopeless. Saying you can't build a general ML model to grade coins from random quality photos is not the same as saying you couldn't build a system to grade specific coins like our hosts.
They already have a set of training photos taken under similar conditions, namely their photo archive. By controlling the imaging process, they eliminate adversarial modifications and many of the issues regarding image quality.
For many coins in many series, they have enough tagged images to train an ML model. Not a system that can grade any arbitrary coin, but one that could grade say an untoned Morgan. This isn't even difficult.
So in one of my industries we can predict accurately within 5% or less with AI. Millions of data points have been loaded. All this will take for AI to run the show is someone to feed the machine the prior millions of human grades and it will calibrate to those grades. Then humans can re-review and kick out the outliers and the machine gets even better. With that said, grading becomes more standardized and more efficient. It’s not far away.
Our hosts don't have those millions of data points.
Since inception they have graded mpre than 75 million collectables, which are not only coins.
They do not have images of them all.
The images that they do have are variable in quality, taken with different equipment and for different purposes
The images are heavily skewed towards certain items (e.g. modern ASEs) which are not representative of collector coins in general.
Can these issues be solved? Sure. But, where is the economics of it? What is in it for me as an owner of several already graded coins to have them photographed? Who pays for it? Who benefits?
Look at the uptake of CoinWorld+ for an instructive example. Where the costs of insured shipping alone have made the whole deal questionable.
-----Burton ANA 50 year/Life Member (now "Emeritus")
While not coin related, I've been very unimpressed with the AI programs in Radiology. Even those looking for the easy "low hanging fruit" are terrible.
Not worried about losing my job to AI - now retired!
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Was at the LB show last week, helping at the Great American Treasures table.
Was offered a Compugrade slab, mid-62 range, ! We weren't able to agree on a reasonable price (to me) so had to pass. He had a second, graded mid-64, but not with him at the show.
Nice to know they are still out there!
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
He wanted $300 - I'm lucky to get $225 on ebay for the common ones.
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Seems like a strong price - haven't seen many offered lately. If I had listed it, $220 max!
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Finally was able to add a Compugrade into my grading set - needed the 64.1 and it had been a long time since something I needed was offered! Took a while to get it, but finally arrived yesterday. I think the seller forgot to ship it!
What looks like a chip on the obv. is the ebay "remove background" flaw. I've never seen a damaged Compugrade slab, that I can recall.
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
I know this was posted in a timely fashion on the US Coin Forum, but want to add it here as I still encounter collectors who don't realize that Astrorat is no longer with us.
At the time, I didn't catch on as to why he sold me so many Compugrade coins, to include his MS 63.0 through 63.9 set.
I still miss "chatting" with him .......
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
There was a toned Morgan offered on ebay recently, the listing ended early since the seller noted holder damage.
It's the 1st damaged Compugrade holder I've ever seen.
Are there other examples that forum members know of and might have pics??
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
@Lakesammman said:
There was a toned Morgan offered on ebay recently, the listing ended early since the seller noted holder damage.
It's the 1st damaged Compugrade holder I've ever seen.
Are there other examples that forum members know of and might have pics??
Did he end up selling it to you? I was bidding and then saw it pulled. I would guess if dropped on a hard surface, that could happen even if it is not common to see damaged Compugrade slabs.
On another note, I think you outbid me on the dmpl...
Well, I tried to buy the damaged holder, but he seemed to think the coin had some value .....
"My friends who see my collection sometimes ask what something costs. I tell them and they are in awe at my stupidity." (Baccaruda, 12/03).I find it hard to believe that he (Trump) rushed to some hotel to meet girls of loose morals, although ours are undoubtedly the best in the world. (Putin 1/17) Gone but not forgotten. IGWT, Speedy, Bear, BigE, HokieFore, John Burns, Russ, TahoeDale, Dahlonega, Astrorat, Stewart Blay, Oldhoopster, Broadstruck, Ricko.
Comments
Posted a couple Saints as a GTG earlier this week - they have been added to the OP in this thread.
Have been tracking the "serial" numbers on the reverse from coins I've seen or own. Here's an example of a higher 1st number with variable following numbers. 1st number unlikely to be a sequential invoice number - could it be a customer ID number??
I bought my Compugrade slab including Compugrade sample slabs over 10 years ago. I currently have a dozen of them. The sample Compugrade were half dollars and they set me back $10 each.
All of the non sample ones are Morgan silver dollars as I recall.
Those Saints are homeruns in the Comnpugrade slabs!
I personally do not think it matters as long as the word COMPUGRADE is in all caps.
I disagree with it helping the hobby. We just have "collectors" that can read numbers on a label, but not able to grade coins. The grading of coins with become a lost art and black magic.
It took me a while to finally be able to post on this thread, but I have something that many people on here will be happy to see........
I still think it's very unlikely.
The kind of AI that is making strides today is machine learning image classification, where we use a set of tagged images to self-program a neural network. This isn't hard, you can do it at home on a Raspberry Pi to differentiate between types of garbage (recyclable or not).
Those models are dependent on having a training set of properly labeled images. The number of images used to train the model impacts the ability of the model to identify a new image.
It might take 10 or 20 images to differentiate between a recyclable can and some food waste. It takes 100s of images of dogs in all orientations, all sizes, all breeds, different image quality, and 100s of cat photos to train the model to differentiate dogs from cats. Interestingly, we can't say how the model works to differentiate the two. But it works.
In the real world, these systems are highly impacted by small changes in the image - there are a number of papers on how tiny modifications of an image (indetectable to a human) lead to misidentification.
Could we program a model to tell us what type of coin something is - certainly. It might take a lot of tagged images if we go back to the era of hand-cut dies, but a basic set of US circulating coins wouldn't be hard. The 50SQ and ATB and Innovation dollar programs up the number of images required, but they aren't hard to get ahold of.
Thus in order for a ML system to grade coins, we need two impossible things:
Yeah right.
Now all is not hopeless. Saying you can't build a general ML model to grade coins from random quality photos is not the same as saying you couldn't build a system to grade specific coins like our hosts.
They already have a set of training photos taken under similar conditions, namely their photo archive. By controlling the imaging process, they eliminate adversarial modifications and many of the issues regarding image quality.
For many coins in many series, they have enough tagged images to train an ML model. Not a system that can grade any arbitrary coin, but one that could grade say an untoned Morgan. This isn't even difficult.
ANA 50 year/Life Member (now "Emeritus")
I think it's a Saint in a Compugrade slab. Tim is going to be happy!
@Lakesammman :-)
ANA 50 year/Life Member (now "Emeritus")
I figured he would be lol!!
That's a tough one to find -- congrats!
Any new Compugrade purchases?? I'm seeing a few now and then on ebay, but haven't found a grade I need for the set in quite a while!
Yep, they seem to show up really occasionally, but the prices are such that I'm not really chasing them. It's not a primary focus and we never really figured out the labels (which is really what I care about).
ANA 50 year/Life Member (now "Emeritus")
John Kraljevich offered a Compugrade brochure on Facebook last night. Due to my ignorance on how auctions work there, I was outbid. John gave me permission to post these pictures.
Did we ever post these? The three Compugrade patents...
https://patents.google.com/patent/US4899392A/en - Method and system for objectively grading and identifying coins
https://patents.google.com/patent/US5133019A/en - Systems and methods for illuminating and evaluating surfaces
https://patents.google.com/patent/US5144495A/en - Systems for illuminating and evaluating surfaces
ANA 50 year/Life Member (now "Emeritus")
Man- those Compugrade ms63 saints look like ms65 in todays time.
proof inflation exists here too!
Image classification as mentioned above would be the answer for identifying what type of coin. However the way classification works would not be very good at grading. It would be identifying features and only those within the training set.
A novel approach that would have promise would be a Siamese neural network. If you train it on images representing wear states I’d say it would do a pretty good job at seeing how similar or not a new object is. And it would have the advantage of when seeing new items it could then be held off and the embedding later labeled and then become part of the comparison set.
TurtleCat Gold Dollars
Actually, as patented - 36 years ago (1987 application) - it was for grading. You are almost looking backwards in time and saying the Wright Brothers should have used GPS to document their first flight.
in the mid 1980s, Neutral networks were a research topic that has basically run its course with the computers of the day and was just being revived as smaller computers became available... When I took CS231 Topics in AI from Jim Low in 1980, NN was just a historical oddity. He did however bring popcorn to the Friday afternoon movie screenings which is why most of us took the course.
[I did a project for Jim implementing (name,type,value) tuple functions in SAIL]
https://cs.stanford.edu/people/eroberts/courses/soco/projects/neural-networks/History/history1.html
https://cs.stanford.edu/people/eroberts/courses/soco/projects/neural-networks/History/history2.html
It's not until the 90s that image processing enters the field and at the time, only with what is today considered toy images (handwriting samples 32x32 pixels of gray scale).
https://towardsdatascience.com/the-brief-history-of-convolutional-neural-networks-45afa1046f7f (paywall, 3 free per month)
Photoshop was introduced in 1990, and 200 K pixels was the best image size out there (for reference 640x480 is 300K pixels).
http://www.digicamhistory.com/1990.html
What is the first paper on Siamese Neural Networks? 1994??? It might be a far better choice than today's throw spaghetti (as many images as you can grab) against the wall and see what sticks... but it's not the sexy stuff that gets the Unicorn valuation for recreating Eliza with more data...
ANA 50 year/Life Member (now "Emeritus")
Hard to say. It was never proven, but widely said that the grading didn't work and the few coins they did slab were human graded. Perhaps with some input from the automation, but...
Certainly the patents are very technical grading ... count the # of dings, weight the counts based on a map of how important the area is to the overall coin, compute a #. And we all know how well technical grading has faired against today's market grading schemes.
The difficulty in creating the maps might indicate why only a very few types of coins have ever been seen in CompuGrade slabs. It was never going to be a general-purpose grading mechanism.
ANA 50 year/Life Member (now "Emeritus")
I'm curious about how that system they built would grade today if you only changed the image acquisition device and added processing power to handle more data. In 1991, they had to use a video camera with and A/D converter to generate what I imagine was a 512x512 image of the coin. That image chain probably had problems with signal overshoot where there were points of very high contrast. Not sure how well they'd have been able to be filtered out. Pull that out and put in a modern DSLR that can produce an image of at least 2048x2048 (16x the information) and see what happens to the grading results. Do they stay the same? Do they consistently shift in one direction? Do they fluctuate a lot? Do they converge more on "the standard"?
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars
On the contrary. My point was not at all about what could have been done or anything like that. It was about what might be a solution to doing it today.
TurtleCat Gold Dollars
I think that's a big issue.
The Compugrade patent references an Image Analysis System that could do 500x magnification as an example of the imaging system that might be used, and says
It will be apparent to those skilled in the art from the following discussion that other types of imaging hardware and/or systems may be utilized in implementing the present invention. For example, scanning electron microscopes, energy dispersive spectrophotometers, VCRs, laser scanners, holography, interferometry and image subtraction are a few of the alternate, presently available types of equipment technologies which may be used.Interestingly, the 2nd & 3rd Compugrade patents were about lighting the images to reveal the surface of the coin...
Even the best imaging system today is nowhere close to 500x lifesize for a full coin. Kodak is quoted as a full frame 35mm image is 6K resolution or something like a 20MP (5568 × 3712 pixels) image. The largest picture ever taken is usually quoted as a NASA 1.5 billion pixel image (69,536 x 22,230).
BTW, a modern mirrorless shoots raw at 30, 45 (8192 x 5464), even 60MP... Yet when DIY images first came to the web, and some people had - gasp - 640x480 cameras, the 'standard' was 320x240. People actually cropped to that, discarding the original image, "because we'll never need higher resolution".
Today we know better... we know imaging books & things at 4k or higher resolution won't suffice forever.
But what does happen when you have to reimage your entire training set? Can you even find the same coins? Must you own them, store them, and plan to re-image them every 5 or 10 years???
What happens when that image isn't raw but is processed through an AI designed to make better people & landscape shots??? Does the camera wait for Miss Liberty to look straight at the lens and not blink? Does it glam up her photo?
What happens when the image sensors change and the color response to even the same lights is different? One of the improvements being worked on is to stack the sensors so you don't have little triangles of RGB sensors but actually read the values at the exact same spot.
ANA 50 year/Life Member (now "Emeritus")
Ah. Run with that... what do we do when the answer isn't to get more labeled images to stuff into the training set? Because the more images, the more people doing the labeling the more disagreements you will have.
And what's true for one series, even for one date range of a series isn't true for others?
We had a preso on Buffalo Nickels two weeks ago at coin club and the presenter was talking about the difficulty in grading the late 20s early 30s coins. We've all seen those gorgeous 1936 and 1937 coins when new equipment was finally able to bring out the details from the dies. But what about say that 1927 coin that never had the details well struck up and you might mistake a MS coin as having the same "look" as a VF 1937.
Do you apply the same standards to 1880/1881 S Morgans (when San Fran was just dialed in and turning out gorgeous coins) to the late New Orleans coins?
ANA 50 year/Life Member (now "Emeritus")
This specific question came up at their presentation 32 years ago, comparing 81-S to 92-O. Their response at the time was that the rules had to be developed for each date and mint individually to take these things into account. Mind you, no AI then like there is today. Simply weights on the degree of deficiencies of the surfaces, luster, and strike.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars
They emphasized the importance of having multiple views of a coin during the presentation, with each view providing specific input that would be used for grading.
The "500x" magnification number is somewhere between silly and meaningless. You see stuff like that on cheap digital microscopes today. Anything can be blown up to an arbitrary magnification. The image might (will) be crap, but at least you have a big number in front of the 'x'.
The bigger issue is that what do you do when your grading scale is rejected by the marketplace? That's a potential fatal bug in your training, so you have to start over with a new training dataset that is perceived as being the most reflective of the current market. Then you end up with grades that change with versions of your software, killing the repeatability claim you touted. The alternative is that you stand fast and nobody uses your service anymore.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars
AI can easily learn to grade in only a matter of hours- and perhaps minutes.
Alpha zero and other such programs proved that
The time for this is now.
Although I disagree with grading a coin via AI, “fingerprinting” for identification is genius.
So in one of my industries we can predict accurately within 5% or less with AI. Millions of data points have been loaded. All this will take for AI to run the show is someone to feed the machine the prior millions of human grades and it will calibrate to those grades. Then humans can re-review and kick out the outliers and the machine gets even better. With that said, grading becomes more standardized and more efficient. It’s not far away.
Our hosts don't have those millions of data points.
Since inception they have graded mpre than 75 million collectables, which are not only coins.
Can these issues be solved? Sure. But, where is the economics of it? What is in it for me as an owner of several already graded coins to have them photographed? Who pays for it? Who benefits?
Look at the uptake of CoinWorld+ for an instructive example. Where the costs of insured shipping alone have made the whole deal questionable.
ANA 50 year/Life Member (now "Emeritus")
While not coin related, I've been very unimpressed with the AI programs in Radiology. Even those looking for the easy "low hanging fruit" are terrible.
Not worried about losing my job to AI - now retired!
Was at the LB show last week, helping at the Great American Treasures table.
Was offered a Compugrade slab, mid-62 range, ! We weren't able to agree on a reasonable price (to me) so had to pass. He had a second, graded mid-64, but not with him at the show.
Nice to know they are still out there!
"Reasonable Price" as in less than the going rate for a kidney?
ANA 50 year/Life Member (now "Emeritus")
He wanted $300 - I'm lucky to get $225 on ebay for the common ones.
Seems like a strong price - haven't seen many offered lately. If I had listed it, $220 max!
Finally was able to add a Compugrade into my grading set - needed the 64.1 and it had been a long time since something I needed was offered! Took a while to get it, but finally arrived yesterday. I think the seller forgot to ship it!
What looks like a chip on the obv. is the ebay "remove background" flaw. I've never seen a damaged Compugrade slab, that I can recall.
Great thread and great project @Lakesammman !
--Severian the Lame
Awesome
ANA 50 year/Life Member (now "Emeritus")
I know this was posted in a timely fashion on the US Coin Forum, but want to add it here as I still encounter collectors who don't realize that Astrorat is no longer with us.
At the time, I didn't catch on as to why he sold me so many Compugrade coins, to include his MS 63.0 through 63.9 set.
I still miss "chatting" with him .......
I miss him too! He was one of the great ones!
ANA 50 year/Life Member (now "Emeritus")
There was a toned Morgan offered on ebay recently, the listing ended early since the seller noted holder damage.
It's the 1st damaged Compugrade holder I've ever seen.
Are there other examples that forum members know of and might have pics??
Did he end up selling it to you? I was bidding and then saw it pulled. I would guess if dropped on a hard surface, that could happen even if it is not common to see damaged Compugrade slabs.
On another note, I think you outbid me on the dmpl...
Well, I tried to buy the damaged holder, but he seemed to think the coin had some value .....