Dynamic 3D Gaussian Fields for Urban Areas

TL;DR

Given a set of heterogeneous input sequences of a common geographic area, we optimize a single dynamic scene representation that permits rendering of arbitrary viewpoints and scene configurations at interactive speeds.

Demo video

We render a novel free-form trajectory across five highly diverse sequences using the same model. We obtain the trajectory by interpolating between keyframes selected throughout a common geographic area at a constant speed of 10 m/s. We render each sequence with its unique apperance and set of dynamic objects, simulating various distinct traffic scenarios.

Abstract

We present an efficient neural 3D scene representation for novel-view synthesis (NVS) in large-scale, dynamic urban areas. Existing works are not well suited for applications like mixed-reality or closed-loop simulation due to their limited visual quality and non-interactive rendering speeds. Recently, rasterization-based approaches have achieved high-quality NVS at impressive speeds. However, these methods are limited to small-scale, homogeneous data, i.e. they cannot handle severe appearance and geometry variations due to weather, season, and lighting and do not scale to larger, dynamic areas with thousands of images. We propose 4DGF, a neural scene representation that scales to large-scale dynamic urban areas, handles heterogeneous input data, and substantially improves rendering speeds. We use 3D Gaussians as an efficient geometry scaffold while relying on neural fields as a compact and flexible appearance model. We integrate scene dynamics via a scene graph at global scale while modeling articulated motions on a local level via deformations. This decomposed approach enables flexible scene composition suitable for real-world applications. In experiments, we surpass the state-of-the-art by over 3 dB in PSNR and more than 100x in rendering speed.

Method

We use sets of 3D Gaussians $G$ as geometry scaffold, neural fields $\phi$ and $\psi$ to represent sequence- and object-specific appearance and geometry variations, and a scene graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ to express the scene configuration at each sequence-time pair $(s, t)$. We condition the neural fields with latent codes $\omega$ of the nodes in $\mathcal{V}$. To render a view at $(s, t)$, we compose the sets of 3D Gaussians using the coordinate system transformations $[\mathbf{R} | \mathbf{t}]$ along the edges $\mathcal{E}$.

Comparison to state-of-the-art

Our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering. We provide qualitative illustrations below. Note that all results are rendered from a single model trained on multiple sequences.

ML-NSG Ours GT
ML-NSG Ours GT
ML-NSG Ours GT
ML-NSG Ours GT
ML-NSG Ours GT
ML-NSG Ours GT

Single-Sequence Results

We show results on the Waymo Open dataset when training on a single sequence as in other contemporary methods. We synthesize novel views along the initial ego-vehicle path in a spiral.

BibTeX

@article{fischer2024dynamic,
    author    = {Fischer, Tobias and Kulhanek, Jonas, and Rota Bul\`{o}, Samuel and Porzi, Lorenzo and Pollefeys, Marc and Kontschieder, Peter},
    title     = {Dynamic 3D Gaussian Fields for Urban Areas},
    journal   = {arXiv preprint arXiv:2406.03175},
    year      = {2024}
}
Tobias Fischer
Tobias Fischer
Ph.D. Student in Computer Vision

Tobias Fischer is a Ph.D. student at the Computer Vision and Geometry Group of ETH Zürich.