Video modification in drone and satellite imagery

Abstract

The ability to create and detect synthetic video is becoming critically important to scene understanding. Techniques for synthetic manipulation and augmentation of data increase diversity within available datasets, while not requiring laborious labeling efforts. That is, the ability to create synthetic video can enable augmentation of small realistic datasets on which to further train Artificial Intelligence and Machine Learning (AI/ML) algorithms. Thus, it may be desirable to add, remove, or modify vehicles in satellite and overhead video. In our previous work, we leveraged Generative Adversarial Networks (GANs) to transform cars into trucks (and vice versa) in static images. We utilized an attention-based masking approach that assists the network in transformation of the object and not background. In addition, we demonstrated the benefits of numerous data augmentation procedures, including presenting a new artificial dataset of vehicles from an aerial perspective and introducing novel augmentation techniques appropriate for our network architectures. This work extends the applied techniques from still imagery to video. We employ a few different architectures: (1) a fully dynamic 3D convolutional discriminator network with static generators, (2) a fully dynamic 3D convolutional discriminator and generator network, and (3) an architecture that computes "warp" between frames for input to a static generator. Additionally, to help enforce consistency, we experiment with an interframe classifier that verifies whether two frames belong to the same video sequence or not. We run experiments on a real-world dataset, presenting promising results in terms of FID, KID, and metrics developed from a classifier trained on our dataset.