3D Gaussian Splatting for Real-Time Radiance Field Rendering

Thu, 9 Nov 2023 19:50:01 +1100

Andrew Pam <xanni [at] glasswings.com.au>

Andrew Pam

"Radiance Field methods have recently revolutionized novel-view synthesis of
scenes captured with multiple photos or videos. However, achieving high visual
quality still requires neural networks that are costly to train and render,
while recent faster methods inevitably trade off speed for quality. For
unbounded and complete scenes (rather than isolated objects) and 1080p
resolution rendering, no current method can achieve real-time display rates.

We introduce three key elements that allow us to achieve state-of-the-art
visual quality while maintaining competitive training times and importantly
allow high-quality real-time (≥ 100 fps) novel-view synthesis at 1080p

First, starting from sparse points produced during camera calibration, we
represent the scene with 3D Gaussians that preserve desirable properties of
continuous volumetric radiance fields for scene optimization while avoiding
unnecessary computation in empty space; Second, we perform interleaved
optimization/density control of the 3D Gaussians, notably optimizing
anisotropic covariance to achieve an accurate representation of the scene;
Third, we develop a fast visibility-aware rendering algorithm that supports
anisotropic splatting and both accelerates training and allows realtime
rendering. We demonstrate state-of-the-art visual quality and real-time
rendering on several established datasets."

Via Wayne Radinsky.

Share and enjoy,
               *** Xanni ***
mailto:xanni@xanadu.net               Andrew Pam
http://xanadu.com.au/                 Chief Scientist, Xanadu
https://glasswings.com.au/            Partner, Glass Wings
https://sericyb.com.au/               Manager, Serious Cybernetics

Comment via email

Home E-Mail Sponsors Index Search About Us