Plateai Is Live For The 1851–1857 3¢ Washingtons (Scott #10, 10A, 11, 11A, 25, 25A — Types I And II)

PlateAI is live for the 1851–1857 3¢ Washingtons (Scott #10, 10A, 11, 11A, 25, 25A — Types I and II).

It's a free, web-based machine-learning tool to plate and study these stamps. Visit the PlateAI tab on stampplating.com.

Even if you don't plate stamps, check your #10s and #10As, the model will tell you the correct Scott number with near-perfect accuracy on any clean copy. You may have something different from what you think.

For platers: upload a scan (1200 DPI is ideal), click Align and Crop, then Plate! The model returns a ranked top-10 list of candidate positions. The correct position is in that top 10 essentially every time. As the top single guess, it's right:

- 97% of the time on 4-margin stamps
- 86% of the time on 3-margin stamps
- 77% of the time when 2 adjacent framelines are gone

PlateAI also includes tools to confirm and study the predictions: side-by-side comparison of your stamp against 4 vetted reference scans, frameline-by-frameline comparison (with aligned compression scans), corner comparison, and a full Position Explorer for browsing any of the 2,600 positions across all 13 plates with or without a patient stamp loaded. You can narrow the search by attributes like type, plate, relief, recuts, and more.

Background: I built this as a way to learn machine learning hands-on, and as a tool to help me plate the stamps I was struggling with. Once I realized the model was genuinely good at the job, I built out the app with other collectors in mind: both 3¢ specialists and the 3¢-curious.

I could not have done it without Robert J. Lampert's patient mentoring on plating this issue, nor J. Bryan O'Doherty's stampplating.com website.

Give it a try. Feedback welcome.

One of the side benefits of this project was the identification of dozens of misplated stamps in the reference collections. Bryan is working on slotting in corrections on the stampplating website. A handful of these had been previously identified, including the Chase stamp.

Collection Orig. Corrected
Amonette 53L3 55L3
Amonette 64R3 29R3
Amonette 20L4 68L7
Amonette 95L4 85L4
Amonette 19L4 57L4
Amonette 28L4 21L4
Amonette 58L7 74L6
Amonette 71L7 52L7
Amonette 82L8 82L7
Amonette 12R8 32R8
Amonette 51L1E 52L1E
Amonette 51R2L 32R2L
Amonette 70L5L 45L2L
Amonette 18R5L 62L5L
Amonette 16L4 34L4
Amonette 90L4 50R4
Amonette 97L6 96L6
Amonette 100L1i 89R1i
Amonette 34R1i 16R1i
LAO 63R0 57R0
LAO 10L3 58L2L
LAO 34L3 24L3
LAO 51L3 61L3
LAO 73L3 79R3
LAO 53R3 53L3
LAO 84R4 85R4
LAO 85R4 94R4
LAO 89R4 49R4
LAO 90R4 50R4
LAO 96R4 92R4
LAO 17L6 18L6
LAO 59L6 66L6
LAO 91L6 91R6
LAO 4L7 4R7
LAO 24R7 25R6
LAO 31R7 71R4
LAO 38L8 7L7
LAO 20R8 58R8
LAO 22R8 49R8
LAO 78R8 25R8
LAO 29L1E 29L1i
LAO 18R1E 2L1E
LAO 22R1E 43R1E
LAO 58R1E 58R1i
LAO 26R1i 22R1i
LAO 56R1i 55R1i
LAO 75L1L 35L1L
LAO 93R2E 57R0
LAO 81L2L 99L2L
LAO 46R2L 46L2L
LAO 55L5E 8R5E
LAO 29R5E 95L5E
LAO 11L5L 21L5L
LAO 65L5L 100R4
LAO 33R5L 18R5L
LAO 48R5L 22R5L
LAO 73R5L 74L5L
LAO 99R5L 92L5L
LAO 100R4 97L4
LAO 87R6 82R6
LAO 53L7 71R7
LAO 29L1E 29L1i
LAO 29R8 79R8
LAO 38L1E 38L1i
LAO 58R1E 58R1i
LAO 78R1E 78R1i
LAO 75L1L 35L1L
LAO 72R1i 57R1i
LAO 80L1L 19L1L
LAO 14L5L 56R5L
Chase 32L2E 55L2E

Wow nice.

Omg. This is absolutely awesome. I just used the PlateAI for about 15 stamps. Everyone one of them, except one, came back with #1 choice as what I had confirmed or marked as the candidate. It even worked really well with less than 1200 dpi images (1000 to 1125 range). I will need to rescan to make it happier and try again.

I went through maybe 30-40 more stamps last night. Outside of a couple of puzzlers, every #1 was a match for a previously confirmed or candidate. Is there a specific place where you want feedback?

Glad it is working for you! Here is ideal for feedback.

I've now plated over 1000 stamps with PlateAI. . .a couple observations. . .on 4-margin stamps, it is very good - almost always right. It is still good with less-than-perfect stamps, but only mostly right with its top choice - this is consistent with testing. I noted two out of about 1000 that did not have a match in the top-20, and I also noted 3-4 stamps that gave a HIGH confidence score on an incorrect plating - these were very poor stamps, but I am not super happy that it claimed high confidence when it was wrong.

This is consistent with the testing I did, but this was a more "real world test" than the 456-stamp "testing panel" of 4-margin stamps with simulated missing framelines through masking. Many of the 1000 I plated through the tool were much worse than just missing framelines. . .and there were only a handful I couldn't plate because too much of the stamp was missing.

The main things though are 1) I plated 1000 stamps in a few hours over a few days 2) The accuracy is perfect or very near perfect (not the AI by itself, but the AI and me together).

If you are into plating or even just confirming the Scott numbers of your 3¢ Washingtons, check it out and let me know if you are having as much success as I am having with it. I see about 40 people have tried it so far and 1200 stamps have been plated besides mine.

I have plated several hundred stamps using the PlateAI with good results. I need mostly non-inner line stamps for my plating so the stamps I've been plating have been (mostly) from plates 4, 6, 7 and 8. So far only 1 decent (3 margin) stamp that Plato failed to find for me and several 2+ margin. It's been finding stamp positions for stamps that I probably would have never tried. This is a great new tool for us platers. Thanks Dave and Bryon.

@philazilla - I'm curious, the plate 4 strip of 3 that you just plated for me - can PlateAI handle multiples, or do they have to be broken down into singles? I presume the latter?

For us human platers, we always crave multiples as they make plating a lot easier. Can the AI leverage this (extra data from multiples) as well, or is it stuck dealing with single stamps ?