VidSfM: Robust and Accurate Structure-From-Motion for Monocular Videos

Area of Science:

• Computer Vision
• Photogrammetry

Background:

• The proliferation of high-quality smartphone videos has increased the demand for large-scale scene reconstruction.
• High-resolution and high-frame-rate videos present challenges like increased match outliers and redundant images, impacting reconstruction accuracy and efficiency.

Purpose of the Study:

• To develop a novel framework for accurate and robust structure-from-motion (SfM) from monocular videos.
• To improve the efficiency and scalability of scene reconstruction systems by utilizing video sequence characteristics.

Main Methods:

• Utilizing spatial and temporal continuity in video sequences to enhance reconstruction accuracy and robustness.
• Employing video redundancy to boost system efficiency and scalability.
• Implementing adaptive loop matching, cluster-based camera registration, local rotation averaging, and local image extension strategies.

Main Results:

• The proposed framework demonstrates superior robustness, accuracy, and scalability compared to existing state-of-the-art methods.
• The system successfully integrates data from multiple video sequences for simultaneous reconstruction.
• Extensive experiments on diverse indoor and outdoor datasets validate the method's performance.

Conclusions:

• The developed SfM framework effectively addresses challenges posed by high-resolution and high-frame-rate monocular videos.
• The approach offers significant improvements in scene reconstruction accuracy, robustness, and scalability.
• The system's ability to integrate multiple video streams enhances its practical applicability.