Message:

<p>[QUOTE=&quot;geekpryde, post: 3912094, member: 36248&quot;]I used to think that you could code a coin grading system based on pictures, or more recently that machine learning could do it.&nbsp;&nbsp; I know machine learning works great for playing chess, and pretty good looking for cancer for example.&nbsp;&nbsp; But the fatal flaw of coin grading from images was pointed out by Doug.&nbsp; The images themselves. </p><p><br /></p><p>From my understanding, you are talking about supplying a single obverse and single reverse photo of each coin.&nbsp;&nbsp; Well, anyone who was spend enough time looking at internet auction photos knows that a specific coin can look different in HA photos, Great Collection photos, PCGS photos, and your own photos.&nbsp; Sure there are coins that look similar across maybe three sets.&nbsp; But I have seen many other examples of the identical coin looking wildly different between sets of images.&nbsp;&nbsp; I am not using that term loosely.&nbsp;&nbsp; Wildly, almost unbelievably different to the point where I did not think they could possibly be the same coin.</p><p><br /></p><p>Here is even a bigger problem with two photos per coin.&nbsp; You are forgetting an entire side! (Spoiler alert: the rim)&nbsp;&nbsp; You need at least 3 photos per coin to even begin grading.&nbsp; But taking into account my first point, I still think it&#039;s futile unless all you care about is pure wear and never need to take into account strike issues.&nbsp;&nbsp; It&#039;s DOA for me from pictures unless your goals are very limited indeed.</p><p><br /></p><p>With videos made under a mechanical method to consistently rotate a coin similar to how a human would do it, under exact and ideal lighting, perfectly capturing all three sides at once, you would have better results.&nbsp;&nbsp; But you could not harvest years of previous photos, so you would be starting at square one.</p><p><br /></p><p>An idea would be to also skip handing the AI the existing TPG grades at all.&nbsp;&nbsp; Give it some goals and let it organize the coins how it sees fit.&nbsp;&nbsp; It might create something really interesting or amazing or weird.&nbsp; &nbsp; Like you could train it by valuing coins with more luster higher than coins with less, which would be higher than coins with none.&nbsp;&nbsp; Value coins with the most detail as higher than less detail higher than no detail.&nbsp;&nbsp; Valuing coins with no contact marks higher than some marks, higher than many marks.&nbsp; Etc.&nbsp;&nbsp; Don&#039;t give it TPG grades, let it create hierarchy of coins based on a limited set of desirable criteria, and see what shakes out.&nbsp; It might blow our minds how it orders the coins.</p><p><br /></p><p>It would be a long and expensive undertaking.&nbsp; Possibly worth the effort as an experiment, or even for just the honor of a featured thread on CT.&nbsp; <img src="styles/default/xenforo/clear.png" class="mceSmilieSprite mceSmilie8" alt=":D" unselectable="on" unselectable="on" />[/QUOTE]</p><p><br /></p>

Your name or email address:

Do you already have an account?

• No, create an account now.
• Yes, my password is:
• Forgot your password?

Stay logged in