AI-POWERED RING SIZE ESTIMATION FOR A REALISTIC VIRTUAL TRY ON

Abstract - This study presents an AI-assisted system to improve ring size estimation and the overall virtual try-on experience. By employing a Convolutional Neural Network (CNN) model and MediaPipe's 21-point hand landmark model, the system is capable of tracking finger size in real-time. By measuring the significant key points of the hand's finger, we can accurately predict the ring size based on Euclidean distances, and then calibrate each measurement against various standard ring sizes. In order to enhance user experience, the virtual try-on experience relies on computer vision and augmented reality (AR). Specifically, OpenCV's Haar Cascade Classifier is used for hand detection, while the wearables are alpha-composited on top of the user's environment. The overlay of the piece and its position is highly realistic and requires minimal setup from the user. This system does not correlate to earlier methods like fixing in an expensive RGB-D camera, nor defined datasets for pre-set landmarks. This virtual try-on experience is unique in how it can combine deep learning with a reduced weight hand detection for fast computation, removing bulkiness while improving accuracy simultaneously. In the future, we plan on implementing further improvements to improve the accuracy of ring size detection, determining support for more jewelry styles, and continuing the transformation of the online jewelry shopping experience.