Methodology for Building Synthetic Datasets with Virtual Humans

摘要: Recent advances in deep learning methods have increased the performance of face detection and recognition systems. The accuracy these models relies on range variation provided training data. Creating a dataset that represents all variations real-world faces is not feasible as control over quality data decreases with size dataset. Repeatability another challenge it possible to exactly recreate ‘real-world’ acquisition conditions outside laboratory. In this work, we explore framework synthetically generate facial be used part toolchain very large datasets high degree environmental variations. Such can for improved, targeted neural networks. particular, make use 3D morphable model rendering multiple 2D images across 100 synthetic identities, providing full image such pose, illumination, background.