TC-SfM: Robust Track-Community-Based Structure-From-Motion.

Area of Science:

• Computer Vision
• Photogrammetry
• 3D Reconstruction

Background:

• Structure-from-Motion (SfM) reconstructs 3D scenes and camera poses from images.
• Ambiguity from visually similar structures leads to inaccurate SfM results.
• Existing methods use additional constraints or two-view geometry to resolve ambiguity.

Purpose of the Study:

• To address reconstruction failures in SfM caused by ambiguous structures.
• To leverage high-level spatial contextual information for improved 3D scene reconstruction.
• To develop a robust method for detecting and correcting ambiguous segments in SfM.

Main Methods:

• Proposed a novel 'track-community' structure to represent scene segments.
• Applied community detection on a track-graph to partition the scene.
• Detected and corrected ambiguous segments using pose consistency checks.
• Aligned partial reconstructions using a bidirectional consistency cost function.

Main Results:

• The track-community approach effectively mitigates reconstruction failures due to visually indistinguishable structures.
• The method accurately merges partial reconstructions from different scene segments.
• Experimental results validate the robustness and accuracy of the proposed SfM technique.

Conclusions:

• Exploiting spatial contextual information via track-communities enhances SfM robustness.
• The proposed method offers a significant improvement over existing techniques for handling structural ambiguities.
• This work provides a reliable solution for accurate 3D scene reconstruction in challenging scenarios.