MBS-NeRF: reconstruction of sharp neural radiance fields from motion-blurred sparse images

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural Radiance Fields (NeRF) enable realistic view synthesis using Multilayer Perceptrons (MLP) for implicit scene representation.
• Degradation in input image quantity and quality hinders NeRF's scene reconstruction and high-quality view synthesis.

Purpose of the Study:

• To develop a NeRF-based framework (MBS-NeRF) capable of reconstructing sharp NeRF from limited, motion-blurred input images.
• To achieve high-quality view synthesis despite input data limitations.

Main Methods:

• Integration of depth information as a constraint to compensate for insufficient view data.
• Introduction of a Motion Blur Simulation Module (MBSM) to model motion blur formation.
• Implementation of camera trajectory optimization during exposure to enhance robustness against incorrect camera positions.
• Training with photometric consistency and depth supervision.

Main Results:

• Successful reconstruction of sharp NeRF from sparse, motion-blurred inputs.
• Demonstrated high-quality view synthesis capabilities validated on synthetic and real datasets.
• Effectiveness of MBS-NeRF in overcoming limitations of degraded input images.

Conclusions:

• MBS-NeRF effectively reconstructs NeRF and synthesizes high-quality views from limited, motion-blurred images.
• The framework addresses key challenges in NeRF reconstruction caused by poor input data quality.
• Depth integration and motion blur simulation are crucial for robust NeRF performance.