Press J to jump to the feed. Press question mark to learn the rest of the keyboard shortcuts
Trust Your (Computers) Eye – Evaluating Round Diamond Cut Quality Using Computer Vision
Until now, my site has relied on the GIA and AGS proportion-based cut grade criteria to identify well-cut round diamonds. Although this helps speed up the search process, Ive always felt it could be improved because not all AGS 0 Ideal proportion combinations show all 8 arrows and have the optical symmetry I typically like to recommend.
Ive developed a computer vision-based method that compares a diamonds image with an image of a canonically ideal diamond to identify diamonds with ideal proportions. Based on experience, an image alone is sufficient to determine ideal proportions, so Ive trained a computer to mimic the judgment of a human eye.
If youre a regular on this sub, then you know Ive heavily criticized the new wave of AI-based diamond recommendation sites. I think they are all marketing and not real AI. Theyre black boxes, lack rigor, and dont validate their model or results to show why they should be believed. They take advantage of a naive public who knows nothing about diamonds or machine learning.
In this post, I go into great detail about my vision-based classifiers performance. I show how my algorithm compares to the AGS 0 Ideal standard and describe how I made trade-offs in terms of the true positive/false positive rate when using it as ground truth. Unlike every other site, I show my models performance so you can evaluate it for yourself – ROC curve, confusion matrix, heatmaps of proportion distributions for the different calls. Maybe you disagree with these algorithmic biases, but at least you know what they are.
I think my algorithm does a better job than going strictly by the AGS 0 Ideal numbers. Based on experience, I think the false positive ideal diamonds my algorithm identifies with 36.5/40.6, 36/40.8, and 34/40.6 crown/pavilion angle combinations tend to show the arrows better than the true positive AGS 0 Ideal diamonds with 41.4 and 41.6 pavilion angles that my algorithm would not identify as ideal based on their image. I would rather have them than the higher pavilion angle AGS Ideals.
I think using computer vision to mimic human vision has resulted in an improvement in search results. The results are live – almost every diamond you click on from James Allen in my price curves will now have excellent optical symmetry and show a strong arrows pattern. Try it for yourself.