US Patent 11188790 Generation of synthetic datasets for machine learning models

Methods are provided for training and validating deep learning models for visual search and related tasks as they pertain to fashion items such as garments. The methods presented address the special needs of visual search as it pertains to fashion and related industries by generating large numbers of synthetic images or videos for training deep learning models, and also for validating those models. Given a 3D model of a target garment, the methods select appropriate 3D models for humans and scenes, and also selects various values for the customizable parameters of each of the 3D models, and then renders an image or video. A dataset comprising such synthetic images can be blended with real-world tagged images to create composite datasets for training.