PSA Grader

Tracks

AI/ML

Inspiration

I recently was going through old card collections and felt inspired to make something that could be used by people debating to send their cards in for PSA grading.

What it does

Analyzes the front of the card's condition and assigns it a PSA grade.

How we built it

• Obtained data through Ebay Developer Program API (~1,000 cards per PSA).
• Filtered out incorrectly labelled titles.
• Cropped images to avoid detection of grades on cards.
• Removed entries that were horizontal, blurry, multi-card listings, etc. (~2,000 cards total)
• Trained a CNN using TensorFlow/Keras.
• Hosted it on a flask application for the user to select photos of their cards and have them graded.

Challenges we ran into

• Getting the data.
• Any type of app development (we are statistics majors, it was tough for us).
• Skewed distributions of training set led to bias in model.

Accomplishments that we're proud of

• Making the model without having experience with visual computing.

What we learned

• I need to work on my understanding of hosting the models I create.
• Quality of training data is important.

What's next for PSA Grader

• More training data for the model.
• More advanced model architectures.
• Case when user inputs a photo that is not a card (it grades high because of training skew).
• Better integration for user experience. (Ability to save collection, see valuations of predicted card value, marketplace).

Built With

• api
• flask
• python
• tensorflow