Efficient Compression Method in Video Reconstruction Using Video Vision Transformer

In training image generators, autoencoders (AEs) are used to compress data into latent variables. However, when attempting to reconstruct video images, learning more complex distributions is necessary, and a simple model does not provide sufficient reconstruction accuracy or require a huge data-set and training time. Therefore, we explore the possibility of using video vision transformer (ViViT) as a video reconstructor. We report the model structure and various settings suitable for using ViViT as an AE in terms of compression ratio and reconstruction loss.