Neural Surfel Reconstruction: Addressing Loop Closure Challenges in Large-Scale 3D Neural Scene Mapping

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• 3D surface reconstruction faces challenges with complex surfaces and scale.
• Implicit neural representations are emerging but struggle with loop closure and bundle adjustment.
• Existing neural methods lack efficient handling of global parameter learning for these tasks.

Purpose of the Study:

• To develop an algorithm for efficient and accurate 3D surface reconstruction.
• To address limitations in loop closure and bundle adjustment for neural reconstruction methods.
• To improve the scalability and reduce the file size of reconstructed models.

Main Methods:

• Leveraging surfels and expanding their definitions for 3D reconstruction.
• Integrating neural descriptors with surfels.
• Framing surfel association as a deformation graph optimization problem.
• Generating a dense volumetric signed distance function (SDF) for mesh reconstruction.

Main Results:

• Effective loop closure detection and correction in challenging scenarios.
• Significant improvement in reconstruction accuracy, with a 16.9% average error reduction.
• Generation of modeling files up to 90% smaller than traditional methods.
• Simplified complexity in 3D neural reconstruction through surfel-level representation.

Conclusions:

• The proposed surfel-based neural method effectively handles loop closure and bundle adjustment.
• This approach offers a more accurate and efficient solution for 3D surface reconstruction.
• The method significantly reduces data storage requirements for reconstructed models.