Sparse-DeRF: Deblurred Neural Radiance Fields From Sparse View

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural Radiance Fields (NeRF) typically require numerous images for accurate scene reconstruction.
• Existing deblurred NeRF (DeRF) methods are impractical for real-world scenarios with few available images.
• Constructing DeRF from sparse views presents significant challenges due to simultaneous blur kernel and NeRF optimization.

Purpose of the Study:

• To develop a method for constructing deblurred neural radiance fields (DeRF) from a limited number of blurry images (sparse-view).
• To address the inherent complexities and overfitting artifacts associated with joint optimization of blur kernels and NeRF from sparse data.
• To enhance the quality and practicality of DeRF in real-world applications where image availability is restricted.

Main Methods:

• Introduced Sparse-DeRF, a novel approach to regularize the joint optimization of blur kernels and NeRF from sparse views.
• Implemented three key regularization components: surface smoothness, modulated gradient scaling, and perceptual distillation.
• Surface smoothness leverages statistical tendencies for accurate scene structure prediction.
• Modulated gradient scaling adjusts backpropagated gradients based on scene object arrangements.
• Perceptual distillation overcomes multi-view inconsistency and compensates for missing clean image information.

Main Results:

• Sparse-DeRF successfully regularizes the complex joint optimization problem.
• Demonstrated alleviated overfitting artifacts and enhanced radiance field quality.
• Achieved effective DeRF construction from as few as 2, 4, and 6 blurry views.
• Extensive quantitative and qualitative experimental results validate the method's effectiveness.

Conclusions:

• Sparse-DeRF offers a practical solution for constructing high-quality deblurred neural radiance fields from sparse-view blurry images.
• The proposed regularization techniques effectively address the challenges of limited multi-view information.
• This work significantly advances the applicability of DeRF in real-world scenarios with image constraints.