MaskViT: Masked Visual Pre-Training for Video Prediction

The paper proposes a pre-training approach called MaskViT for video prediction, based on two design decisions - spatial and spatiotemporal window attention and variable percentage mask ratio. It outperforms prior works in video prediction, can generate high-resolution videos, and helps in planning on a real robot.

Businesses can leverage this approach for video prediction tasks in complex environments, enabling better planning solutions. It can be used to improve the efficiency of operations and workflows, and lead to better decision-making.