I assume you're planning to use these for a machine learning problem. If so, you may have already thought of all of this, but it might be worth pointing out that ensuring you get flips with non-standard descriptions and grades make it into your data sample (e.g., look for flips with 4 rows of text, not just 2, altered slabs, and cards from all grades) might make the model more robust. It also might perform better against your out-of-sample data (cards that people later upload) if your testing data set includes images from different sources with differing degrees of contrast and image resolutions. You may also want to think about using both photos and scans as opposed to just scans if you anticipate users uploading both at some point. I think BGS flips could prove to be the most challenging because of how they reflect light.
|