Core Concepts
RING-NeRF introduces inductive biases for versatile and efficient neural fields, enabling adaptive reconstruction and robustness.
Abstract
Introduction:
RING-NeRF proposes a simple architecture with inductive biases for Neural Radiance Fields (NeRF).
NeRF Representation:
NeRF uses MLP to represent 3D scenes, inspiring many works but facing convergence issues due to large networks.
Current Solutions:
Most solutions focus on scene nature, robustness, and extensibility limitations of NeRF architectures.
RING-NeRF Architecture:
Introduces continuous multi-scale representation and decoder's latent space invariance for improved performance.
Distance-Aware Mapping:
Utilizes distance-aware forward mapping to adjust LOD based on observation distance for accurate renderings.
Coarse-To-Fine Optimization:
Implements continuous coarse-to-fine optimization for stable reconstruction and LOD extensibility.
Experiments:
Evaluates RING-NeRF on novel view synthesis, anti-aliasing, few viewpoints supervision, SDF reconstruction without initialization, and LOD extensibility.
Stats
PSNR values and training times are inset.
PSNR ↑ SSIM ↑ LPIPS ↓ PSNR ↑ SSIM ↑ LPIPS ↓ PSNR ↑ SSIM ↑ LPIPS ↓ PSNR ↑ SSIM ↑ LPIPS ↓ Time ↓
24.62 0.778 0.347 17.61 1.4h
30.79 0.916 0.0761 13.69 3.40 h
37.18 0.969 0.0194 5.71 1.28 h
Quotes
"RING-NeRF has the distinctive ability to dynamically increase the resolution."
"Our contributions include an architecture that represents scenes with continuous detail."
"RING-NeRF outperforms state-of-the-art solutions in terms of quality-speed trade-off."