Please Before Reaching Out About the Grader Not Working Or Something Of That Nature
THIS WAS NOT TRAINED ON POKEMON CARDS OR YUGIOH THE CARD SARE SHAPED AND SIZED DIFFERENT AND THE PRINTING COLOR PATTERNS ARE DIFFERENT . pLUS WITH TH ECARD SBEING ARTISTIC NOT PHOTOS. THE MODEL STILL HAS ISSUES WITH SOME OF THE KABOOM CARDS. iM SUR EIT WORKS TO A POINT BUT IT WAS TRAINED ON 300,000 SPORTS TRADING CARDS.
When You Upload Your Card Image
the 4 bulleted items are on the upload page fpr your use
Crop your picture so it leaves only a slight border outside the card.
Make sure your picture is clear and well lit for best results.
This is super important, blurry pics, the model sees a bad card
Crooked pics the model can read it as trimmed or not the right shape
Make sure to use a good clean pic of the card, with good color and lighting, without
altering or bluring the card for the best accurate grade.
Because I Assure You It Works and Works Well
Model Stats
Dataset = Dataset was over 300000 priofessionally graded cards
The model was trained on over 70 iterations of the 300000 card dataset
The model was trained to ann accuracy of 88.3% with a val-acc of 85.6%
The model ws trained to a loss of 26.4% and a val-loss of 29.3%
Dataset = then due to the model learning to read the labels it was scraped and retrained
This time on graded cards all cropped dwith no labels and no case and suffled with non descript filenames
The model was trained to ann accuracy of 77.1% with a val-acc of 83.7%
The model ws trained to a loss of 36.8% and a val-loss of 39.2%