Video Gaussian Masked Autoencoders for Video Tracking

This paper introduces a novel self-supervised approach, Video-GMAE, for video representation learning. The core idea is to represent a video as a set of 3D Gaussian splats that move over time. This inductive bias allows the model to learn meaningful representations and achieve impressive zero-shot tracking performance. The significant performance gains on Kinetics and Kubric datasets highlight the effectiveness of the proposed method.