Trifocal Relative Pose From Lines at Points.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Relative camera pose estimation is crucial for 3D reconstruction and scene understanding.
• Existing methods struggle with minimal problems involving point-line correspondences in multiple views.
• Gröbner basis methods are computationally intensive for these complex geometric constraints.

Purpose of the Study:

• To develop an efficient solver for two minimal problems in three-view relative camera pose estimation.
• To address cases involving point and line correspondences that are intractable for current state-of-the-art techniques.
• To enhance the robustness and applicability of Structure from Motion (SfM) initialization.

Main Methods:

• Introduced a novel Homotopy Continuation (HC) solver framework named MINUS.
• Specialized HC methods to generic cases of three-view minimal problems (three points/one line, three points/two lines).
• Characterized the number of solutions for these minimal problems.

Main Results:

• The MINUS framework significantly accelerates HC solving for camera pose estimation.
• Simulated experiments demonstrate numerical robustness and stability under image noise.
• Real-world experiments confirm the effectiveness of SIFT features for point-and-line correspondences in three-view reconstruction.
• Successfully solved challenging cases with sparse or noisy matches where traditional SfM initialization fails.

Conclusions:

• The proposed HC solver provides an efficient and robust solution for challenging three-view relative camera pose estimation problems.
• This method expands the capabilities of SfM, particularly in scenarios with limited or noisy feature data.
• The findings contribute to more reliable and accurate 3D scene reconstruction from multiple viewpoints.