Similarity Graph-Based Camera Tracking for Effective 3D Geometry Reconstruction with Mobile RGB-D Camera

Area of Science:

• Computer Vision
• 3D Reconstruction
• Robotics

Background:

• Real-time 3D reconstruction from RGB-D sensors is crucial for applications like augmented reality and robotics.
• Existing online methods often struggle with error-prone frames in the input stream, degrading reconstruction quality.
• Consumer-grade mobile sensors offer accessibility but present challenges in data quality and processing.

Purpose of the Study:

• To develop a novel, robust 3D reconstruction algorithm for mobile RGB-D sensors.
• To improve the accuracy and reliability of 3D reconstruction by intelligently selecting and ordering input frames.
• To automatically filter out error-prone images, enhancing the overall reconstruction quality.

Main Methods:

• Organizing input RGB-D frames into a weighted similarity graph.
• Finding a maximum spanning tree within the graph to determine optimal frame processing order.
• Extending the algorithm with local tree repair and merging disconnected components for enhanced reconstruction.

Main Results:

• Demonstrated robust 3D reconstruction by processing frames in a carefully selected order, rather than strict acquisition order.
• Successfully filtered out error-prone images, leading to a cleaner and more accurate 3D output stream.
• The proposed method enhances existing state-of-the-art 3D reconstruction techniques.

Conclusions:

• A novel frame selection and ordering strategy significantly improves 3D reconstruction from mobile RGB-D data.
• The algorithm effectively handles noisy input, providing a more reliable stream for downstream applications.
• This approach offers a valuable preprocessing step for advanced 3D reconstruction pipelines.