Mega-NeRF: Scalable Construction of Large-Scale NeRFs for Virtual Fly-Throughs (2112.10703v2)

Abstract: We use neural radiance fields (NeRFs) to build interactive 3D environments from large-scale visual captures spanning buildings or even multiple city blocks collected primarily from drones. In contrast to single object scenes (on which NeRFs are traditionally evaluated), our scale poses multiple challenges including (1) the need to model thousands of images with varying lighting conditions, each of which capture only a small subset of the scene, (2) prohibitively large model capacities that make it infeasible to train on a single GPU, and (3) significant challenges for fast rendering that would enable interactive fly-throughs. To address these challenges, we begin by analyzing visibility statistics for large-scale scenes, motivating a sparse network structure where parameters are specialized to different regions of the scene. We introduce a simple geometric clustering algorithm for data parallelism that partitions training images (or rather pixels) into different NeRF submodules that can be trained in parallel. We evaluate our approach on existing datasets (Quad 6k and UrbanScene3D) as well as against our own drone footage, improving training speed by 3x and PSNR by 12%. We also evaluate recent NeRF fast renderers on top of Mega-NeRF and introduce a novel method that exploits temporal coherence. Our technique achieves a 40x speedup over conventional NeRF rendering while remaining within 0.8 db in PSNR quality, exceeding the fidelity of existing fast renderers.

• The paper introduces Mega-NeRF, which scales NeRF to large scenes using modular partitioning and geometric clustering for faster training and interactive rendering.
• It employs a sparse network structure that trains independent submodules on curated data, achieving a 3x training speed improvement and a 12% boost in PSNR.
• The fast rendering approach leverages temporal coherence to deliver a 40x speed-up, enabling real-time virtual fly-throughs in complex urban environments.

Scalable Training and Rendering of Large-Scale NeRFs for Interactive 3D Visual Explorations

Introduction

Neural Radiance Fields (NeRFs) have shown significant promise in creating photo-realistic 3D environments from 2D images. However, adapting NeRF to large-scale scenes such as city blocks introduces challenges, including managing massive datasets with varying lighting conditions, prohibitive model capacities, and the need for fast rendering to enable interactive experiences. The paper presents Mega-NeRF, which addresses these issues through a modular approach that efficiently scales NeRF to unprecedented scene sizes.

Approach

Model Architecture

Mega-NeRF introduces a sparse network structure optimized for large-scale applications, leveraging spatially-aware partitioning of training data to individual NeRF submodules, allowing for parallelized model training. This setup not only reduces training time but also enhances rendering flexibility, enabling near-interactive exploration speeds for virtual fly-throughs in massive environments.

Training Process

A novel aspect of Mega-NeRF is its geometric clustering algorithm, which partitions training images based on visibility statistics to relevant NeRF submodules. Each submodule is trained independently on a curated subset of the data, achieving significant efficiency gains. The training refinement iteratively improves by focusing on scene areas with higher detail requirements, avoiding wasteful computation on less complex regions.

Interactive Rendering

The paper also debuts a fast rendering approach tailored to Mega-NeRF's modular architecture. Leveraging temporal coherence, the rendering technique recycles previously computed scene information, substantially accelerating frame generation while maintaining high fidelity. This method offers a pragmatic balance between rendering speed and quality, crucial for interactive applications.

Experiments and Results

Evaluated on multiple datasets, including novel drone-captured scenes, Mega-NeRF demonstrates a 3x improvement in training speed and a 12% increase in PSNR compared to existing methods. Furthermore, the proposed rendering technique achieves a 40x speed-up over traditional NeRF rendering, with minimal quality loss, indicating its efficacy for real-time applications.

Implications and Future Work

Practically, Mega-NeRF enables the fast creation and navigation of high-fidelity 3D models from large-scale visual captures, a significant advancement for use cases like urban planning and virtual tourism. Theoretically, it pushes the understanding of how to efficiently structure and process neural radiance fields across vast spaces.

On the horizon, combining Mega-NeRF with emerging techniques in dynamic scene handling and more sophisticated machine learning models could further enhance its versatility. Continuous improvements in training and rendering efficiencies will likely open new avenues for NeRF applications, potentially extending to real-time interactive systems on consumer-grade hardware.

Conclusion

Mega-NeRF represents a substantial step forward in the scalability of Neural Radiance Fields, facilitating the practical use of this promising technology in vast, complex environments. Through innovative modifications to traditional NeRF architectures and processes, it offers a path toward seamlessly bridging the gap between detailed 3D scene reconstruction and the dynamic, interactive exploration of such environments.