Zero-Pose-Prior NeRF: Recursive Radiance Field Reconstruction From Unposed and Unordered Images.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural Radiance Fields (NeRF) require accurate camera poses for scene reconstruction.
• Existing methods for pose-free NeRF struggle with complex scenes and large camera motions.
• Prior knowledge or reasonable initialization is often necessary for current pose-free approaches.

Purpose of the Study:

• To develop a novel method for recovering radiance fields from unposed and unordered image collections without prior knowledge.
• To address the critical obstacle of camera pose dependence in NeRF applications.
• To enable robust scene reconstruction and high-fidelity view synthesis from challenging datasets.

Main Methods:

• Proposed Zero-Pose-Prior NeRF, decomposing the problem into self-bootstrapping sub-problems.
• Implemented scene partitioning for hierarchical structure and local-to-global registration order.
• Devised conditionally-decoupled positional encoding for pose estimation and scene representation.
• Developed recursive registration for estimating local poses and unifying them into a global pose space.

Main Results:

• Achieved accurate camera pose estimation without any prior knowledge.
• Demonstrated robust radiance field reconstruction from unposed and unordered images.
• Outperformed state-of-the-art pose-free methods in experiments on real-world scenes.
• Enabled high-fidelity view synthesis, showcasing improved scene representation.

Conclusions:

• Zero-Pose-Prior NeRF effectively reconstructs radiance fields from unposed image collections.
• The proposed method overcomes limitations of existing NeRF techniques in complex scenarios.
• This approach significantly advances the potential for real-world applications of NeRF without pose priors.