Fast Learning Radiance Fields by Shooting Much Fewer Rays.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Radiance fields enable impressive novel view synthesis but suffer from lengthy training times.
• Existing methods to accelerate training are often incompatible with many radiance field approaches.
• A general, efficient training acceleration strategy is needed for radiance field methods.

Purpose of the Study:

• To develop a universally applicable method for significantly reducing the training time of radiance field models.
• To maintain or minimally impact the accuracy of learned radiance fields despite accelerated training.

Main Methods:

• Implemented a strategy to reduce ray shooting during multi-view volume rendering, focusing on pixels with high color variance.
• Introduced adaptive quadtree subdivision of views to dynamically allocate more rays to complex regions with higher rendering error.
• Evaluated the proposed method across various radiance field techniques and standard benchmarks.

Main Results:

• Achieved substantial reductions in training burden by shooting fewer rays, particularly at pixels with dramatic color changes.
• Demonstrated that the method causes negligible impact on the accuracy of learned radiance fields.
• Showcased comparable accuracy to state-of-the-art methods with significantly faster training times.

Conclusions:

• The proposed general strategy effectively accelerates radiance field training across diverse methods.
• Selective ray casting and adaptive subdivision offer a robust approach to optimize rendering efficiency.
• This method presents a practical solution for faster development and deployment of radiance field applications.