Add to Chrome
Add to Firefox
Add to EdgeCity-on-Web: Real-time Neural Rendering of Large-scale Scenes on the Web
Dec 27, 2023
NeRF has significantly advanced 3D scene reconstruction, capturing intricate details across various environments. Existing methods have successfully leveraged radiance field baking to facilitate real-time rendering of small scenes. However, when applied to large-scale scenes, these techniques encounter significant challenges, struggling to provide a seamless real-time experience due to limited resources in computation, memory, and bandwidth. In this paper, we propose City-on-Web, which represents the whole scene by partitioning it into manageable blocks, each with its own Level-of-Detail, ensuring high fidelity, efficient memory management and fast rendering. Meanwhile, we carefully design the training and inference process such that the final rendering result on web is consistent with training. Thanks to our novel representation and carefully designed training/inference process, we are the first to achieve real-time rendering of large-scale scenes in resource-constrained environments. Extensive experimental results demonstrate that our method facilitates real-time rendering of large-scale scenes on a web platform, achieving 32FPS at 1080P resolution with an RTX 3060 GPU, while simultaneously achieving a quality that closely rivals that of state-of-the-art methods. Project page: https://ustc3dv.github.io/City-on-Web/
