Learning Spherical Radiance Field for Efficient 360° Unbounded Novel View Synthesis.

Area of Science:

• Computer Vision
• Computer Graphics

Background:

• Neural Radiance Fields (NeRF) excel at novel view synthesis for bounded scenes.
• Existing NeRF methods struggle with 360° unbounded scenes due to unconstrained camera poses and depth variations.

Purpose of the Study:

• To develop an efficient method for novel view synthesis in 360° unbounded scenes.
• To address the limitations of current neural rendering techniques in large-scale, unconstrained environments.

Main Methods:

• Representing 3D scenes using multiple concentric spheres with implicit layered scene encoding.
• Parameterizing spheres with equirectangular projection images and using a shallow MLP for density and color inference.
• Employing an occupancy grid for accelerated ray sampling and reduced computation.

Main Results:

• Achieved state-of-the-art results on three benchmark datasets for 360° unbounded scenes.
• Demonstrated significant speedup (400x) compared to Mip-NeRF 360 with under 30 minutes of training.
• Showcased competitive accuracy and efficiency on a bounded dataset.

Conclusions:

• Spherical Radiance Fields (SRF) offer an effective and efficient solution for novel view synthesis in 360° unbounded environments.
• The proposed occupancy grid significantly enhances training and rendering speed.
• SRF provides a robust alternative for both bounded and unbounded scene rendering.