Pix2Pt Map for Transfer-Based Few-Shot Learning

Following the rapid increase of vehicles in the streets, it has become a daunting task to automatically identify tire patterns to aid with traffic accident management. Hence, this paper aims to provide a high accuracy tire classification system using Meta-Learning. We use a transfer-based method known as “Few-Shot learning with leveraged feature distribution method” (PT + MAP), upgrading it with an image segmentation model, pix2pix (‘unet128’), to pre-process the image before feeding it into the PT+MAP.The current approach utilizes three steps to build the method:1)Pre-process Tire patterns image as a Mask (Image preprocessing).2)Using a pre-trained backbone to obtain a Gaussian-like distribution by pre-processing the feature vector extracted3)Using the optimal transport-inspired algorithm to leverage the pre-processing and perform prediction.Our upgraded architecture results in high accuracy for tire classification and hence has shown our system’s applicability as an effective automatic tire identification system.