Scaling 4D Representations

TL;DR

This paper demonstrates that self-supervised learning with masked auto-encoding on large-scale video datasets effectively scales to improve performance on non-semantic 4D vision tasks like pose estimation and tracking, with models up to 22B parameters.

Contribution

It shows that scaling transformer-based video models with self-supervised learning improves performance on 4D spatial-temporal tasks, extending the benefits of scaling beyond semantic tasks.

Findings

Performance improves with model size from 20M to 22B parameters.

Scaling benefits are consistent across various 4D tasks.

Large models outperform previous state-of-the-art on non-semantic video tasks.

Abstract

Scaling has not yet been convincingly demonstrated for pure self-supervised learning from video. However, prior work has focused evaluations on semantic-related tasks $\unicode{x2013}$ action classification, ImageNet classification, etc. In this paper we focus on evaluating self-supervised learning on non-semantic vision tasks that are more spatial (3D) and temporal (+1D = 4D), such as camera pose estimation, point and object tracking, and depth estimation. We show that by learning from very large video datasets, masked auto-encoding (MAE) with transformer video models actually scales, consistently improving performance on these 4D tasks, as model size increases from 20M all the way to the largest by far reported self-supervised video model $\unicode{x2013}$ 22B parameters. Rigorous apples-to-apples comparison with many recent image and video models demonstrates the benefits of scaling 4D representations. Pretrained models are available at https://github.com/google-deepmind/representations4d .