Abstract

Visual understanding of the world goes beyond the semantics and flat structure of individual images. In this work, we aim to capture both the 3D structure and dynamics of real-world scenes from monocular real-world videos. Our Dynamic Scene Transformer (DyST) model leverages recent work in neural scene representation to learn a latent decomposition of monocular real-world videos into scene content, per-view scene dynamics, and camera pose. This separation is achieved through a novel co-training scheme on monocular videos and our new synthetic dataset DySO. DyST learns tangible latent representations for dynamic scenes that enable view generation with separate control over the camera and the content of the scene.

Model

Model Overview

The Dynamic Scene Transformer (DyST) learns a decomposed representation of a dynamic scene into scene content, camera pose and dynamics that allows for controllable video generation. DyST is trained by synthesizing novel views using the camera and dynamics control latents. By estimating these control latents from views with matching camera and dynamics on synthetic data, a separation between camera and dynamics is induced. This separation is then transferred to real world video through a co-training scheme. At test time, DyST allows video generation with separate control over camera and content from a few input frames of a real video.

Examples and Visualizations

PCA Analysis and Latent Manipulations (DySO dataset)
Video Manipulations

DySO Dataset

The release of the DySO dataset is in progress.

Related Projects

Scene Representation Transformer (SRT)
Really Unposed SRT (RUST)

Reference

  @article{seitzer2023dyst,
    author = {
      Seitzer, Maximilian
      and van Steenkiste, Sjoerd
      and Kipf, Thomas
      and Greff, Klaus
      and Sajjadi, Mehdi S. M.
    },
    title = {{DyST: Towards Dynamic Neural Scene Representations on Real-World Videos}},
    journal = {arXiv:2310.06020},
    year = {2023},
	url = {https://arxiv.org/abs/2310.06020}
  }