Cascaded and Generalizable Neural Radiance Fields for Fast View Synthesis

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Generalizable view synthesis methods offer high-quality novel views but suffer from slow rendering speeds.
• Scene-specific methods provide efficient rendering but lack generalization to unseen data.
• Neural radiance fields (NeRFs) are computationally intensive due to uniform point sampling.

Purpose of the Study:

• To develop a novel method for fast and generalizable view synthesis.
• To address the limitations of slow rendering in generalizing methods and lack of generalization in scene-specific methods.
• To achieve efficient and accurate novel view rendering using neural radiance fields.

Main Methods:

• Proposed CG-NeRF, a cascade and generalizable neural radiance fields method.
• Introduced a coarse radiance fields predictor and a convolutional-based neural renderer.
• Inferred consistent scene geometry using implicit neural fields and rendered new views efficiently on a single GPU.

Main Results:

• Trained CG-NeRF on the DTU dataset, demonstrating high-quality and accurate novel view synthesis on unseen real and synthetic data.
• Achieved high-speed rendering on a single GPU without additional explicit representations.
• Outperformed state-of-the-art generalizable neural rendering methods on various datasets.

Conclusions:

• CG-NeRF successfully combines fast rendering with generalization capabilities for view synthesis.
• The proposed architecture effectively infers scene geometry and renders novel views efficiently.
• CG-NeRF represents a significant advancement in neural rendering for computer vision and graphics applications.