CVGN: Conditional Image/Video Generation Network with Generative Adversarial Network

An automatic, on-demand software solution is desired that is capable of generating large scale image/video training datasets for downstream analytic algorithms such as object classification or target tracking, with user selected attributes such as types of foreground objects, background scenes, weather conditions, camera view angles, etc., given a small labeled seed dataset or even no seed data at all. To address this critical need and limitations of state-of-the-art data generation approaches, Intelligent Automation, Inc. (IAI) proposes to design and implement an automatic, on-demand Generative Adversarial Network (GAN) based algorithm framework called CVGN, for large dataset generation that can be tailored to DoD’s special needs, e.g., generating infrared (IR) images and videos from overhead camera angles such as Unmanned Aerial System (UAS). The key innovation of the proposed approach is seamlessly combining unsupervised representation learning, a high fidelity simulation engine, transfer learning, and data augmentation with on-demand training data generation.