Abstract

Recent techniques for real-time view synthesis have rapidly advanced in fidelity and speed, and modern methods are capable of rendering near-photorealistic scenes at interactive frame rates. At the same time, a tension has arisen between explicit scene representations amenable to rasterization and neural fields built on ray marching, with state-of-the-art instances of the latter surpassing the former in quality while being prohibitively expensive for real-time applications. In this work, we introduce SMERF, a view synthesis approach that achieves state-of-the-art accuracy among real-time methods on large scenes with footprints up to 300 m^2 at a volumetric resolution of 3.5 mm^3. Our method is built upon two primary contributions: a hierarchical model partitioning scheme, which increases model capacity while constraining compute and memory consumption, and a distillation training strategy that simultaneously yields high fidelity and internal consistency. Our approach enables full six degrees of freedom (6DOF) navigation within a web browser and renders in real-time on commodity smartphones and laptops. Extensive experiments show that our method exceeds the current state-of-the-art in real-time novel view synthesis by 0.78 dB on standard benchmarks and 1.78 dB on large scenes, renders frames three orders of magnitude faster than state-of-the-art radiance field models, and achieves real-time performance across a wide variety of commodity devices, including smartphones.

Video

Real-Time Interactive Viewer Demos

How we boost representation power to handle large scenes

(a): We model large multi-room scenes with a number of independent submodels, each of which is assigned to a different region of the scene. During rendering the submodel is picked based on camera origin. (b): To model complex view-dependent effects, within each submodel we additionally instantiate grid-aligned copies of deferred MLP parameters \(\theta\). These parameters are trilinearly interpolated based on camera origin \(\mathbf{o}\). (c): While each submodel represents the entire scene, only the submodel's assiociated grid cell is modelled with high resolution, which is realized by contracting the submodel-specific local coordinates.

Getting the maximum out of our representation via distillation

We demonstrate that image fidelity can be greatly boosted via distillation. We first train a state-of-the-art offline radiance field (Zip-NeRF). We then use the RGB color predictions \(\mathbf{c}\) of this teacher model as supervision for our own model. Additionally, we access the volumetric density values \(\tau\) of the pre-trained teacher by minimizing the discrepancy of volume rendering weights between teacher and student.

Citation

If you want to cite our work, please use:

@misc{duckworth2023smerf,
      title={SMERF: Streamable Memory Efficient Radiance Fields for Real-Time Large-Scene Exploration},
      author={Daniel Duckworth and Peter Hedman and Christian Reiser and Peter Zhizhin and Jean-François Thibert and Mario Lučić and Richard Szeliski and Jonathan T. Barron},
      year={2023},
      eprint={2312.07541},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}
                

Acknowledgements

The website template was borrowed from Michaël Gharbi. Image sliders are based on dics.