Rock Paper Scissors Dataset

The Rock-Paper-Scissors image classification is a common task for beginners learning to train their own CNN models. Along with using pre-existing datasets, I included my own pictures, specifically focusing on capturing images of hands shown with the palm facing forward, like when you naturally raise your hand in front of a webcam.

Dataset specs:

• Contents: about 2700 pictures of my own, my wife’s and my son’s hands (10 years old).
• Coverage: Both sides of the hand as well as side pictures when rotating the hand
• Background: Photos were taken on a grey uniform background with different lighting conditions
• Organisation: Pictures are ordered in 3 classes directory: rock, paper and scissors
• File format: PNGs, 300×300 px
• Naming convention: [class_id.png]. Example: rock_5.png, paper_258.png

In my GitHub repository for this project, you’ll find notebooks to get the dataset ready and train it. I’ve also included tools to remove backgrounds from images. The final project has a capture tool I developed to take these pictures.