AI Shows Where It Looks On The Stamp To Identify It. (Franklin)

Hey all. The AI model creates a heat map to show where it is looking to identify these stamps. Check it out. It was wrong on 596 and many others but it is starting to get 632 almost every single time.

What colors are reperforated and regummed?

How does it, or you, determine the proper heat-focus location? On some issues, the center of the stamp should quickly determine the catalog number. On others, the border and perforation rate might be better. And, with training, I bet it could determine reperforations with a high degree of probability.

Regumming would need to look at the back, and possibly at an oblique angle, of course.

How would AI deal with differentiating rotary press stamps from flat press stamps when all other potential variables such as perforation rate are identical? The obvious way would be design size measurements but wouldn't that require certain image parameters being met? I am very curious about the current limits of the technology and where it still is more practical to use traditional methods.

If I can learn to tell them all apart, then anyone can. All it takes is study and a desire to learn. I trust my ability more than I trust this. Which takes all the fun out of collecting anyway.

Exactly

What AI program is this?

Hey Guys,

So Stampdoc, this is my program and AI model, Rev collector it still has difficulties identifying the ones we have difficulty identifying and makes a lot of mistakes right now. Rogdcam and partime , about the rotary press flat press, It doesn't identify that, like partime suggested it is a trained thing. There are two sets of folders, a validation set and and a training set of images and the AI model "learns" the category by testing the training set against the validation set. So the more it is trained partime the better it gets at identifying the stamps and yes, I wanted to know where it was looking and it is somewhat interesting to see that without any knowledge of the identification history we have, its looking at vertical perforations, sometimes its looking at the black space, sometimes the corners...

Identifying stamps is something AI could get very good at. Even a general-purpose LLMs like ChatGPT is already useful in simple situations. AI could also plate stamps and go well beyond identifying the catalog number. And I totally disagree with revcollector and rogdcam - this new technology adds to the fun!

Whether we like it or not, this is a glimpse of the future of stamp collecting.

imodius, do you plan to make your software open-source or closed-source?

On occasion, I've used an app called Stamp Identifier. You point your phone camera at a stamp, and it shows you its best guess. Usually pretty accurate, but sometimes I have no idea where it went. I'm sure there is some rudimentary AI behind that app. You may want to check it out for comparison.

Saying "it still has difficulty identifying stamps we all have trouble with" is disparaging collectors who can identify difficult stamps. And if you have fun being told what you have rather then figuring it out for yourself, so be it. Assuming that it's actually correct. Which you will not know. I would certainly rather be able to tell accurately for myself.

Again, exactly this. It doesn't work for structural engineering, medicine or self-driving vehicles. In the end a human needs to verify and not trust blindly. If you want to turn stamp collecting into a production process that is a different animal entirely. Not relaxing to me but to each their own.

Using AI != "trusting blindly"

AI has its advantages. Earlier this week, research was published that showed an AI application was much better and more accurate than specialists at diagnosing lung cancer in a very early stage. Apparently, AI is better at considering all variables, calculating probabilities, and learning from its mistakes than specialists are.

Still, on another forum, someone asked ChatGPT to identify three Finnish stamps. The OP was impressed, but ChatGPT did not much more than assigning three possible FACIT numbers in ascending order to the three stamps from left to right. It, further, could not differentiate between Ojakkala and Aland, and came up with a lot of speculative garbage if not utter nonsense. It further could not link the Facit numbers to the correct series or even years of issue.

From one of the posts about the same Finnish stamp here, I doubt it got the catalogue numbers correct as the lowest number does not appear to be among the posted stamps. It is likelier at least two stamps were the same but for the "shade" of green.

Sadly, people post the garbage on stamp forums, thus disseminating the garbage and helping ChatGPT believe its own nonsense because it can find it posted on the web.

I do not see the problem with using AI to suggest candidate stamps. But you should not blindly believe AI's suggestions.

Take the "heat maps" shown in this thread. AI is not coming up with these areas by its brilliant self. It was told where to look which brings up more questions than it answers.