Camera Pose Fusion via Geometric and CNN Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for relative camera pose estimation from two perspective views in Structure-from-Motion (SfM) face challenges such as degraded performance under scarcity of correspondences, degeneracy in geometric configurations, and difficulties in distinguishing rotational and translational flows.
Innovation Solution
A computer-implemented method that fuses geometrical and Convolutional Neural Network (CNN) relative camera pose estimates by inputting two images into both a geometric solver branch and a CNN branch to obtain initial pose solutions and uncertainty values. These solutions are then fused probabilistically using Bayes' rule to obtain a fused pose.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If geometric solver is used for relative camera pose estimation, then the method provides theoretical foundation and works well under ideal conditions, but performance degrades under scarcity of correspondences and geometric degeneracy
Solution Approach 1:
The patent combines geometric solver and CNN-based estimator into a unified framework where both methods process the same input images and their outputs are fused. The geometric solver provides theoretically grounded pose estimates while the CNN provides robustness under degenerate conditions, and their combination resolves the contradiction between reliability under ideal conditions and precision under challenging conditions.
Solution Approach 2:
The patent creates a composite estimation system that integrates two different estimation approaches (geometric and deep learning) with complementary strengths. The geometric component ensures theoretical correctness while the CNN component adds robustness, creating a composite solution that maintains both reliability and precision across varying conditions.
2Reliability
If CNN-based estimator is used for relative camera pose estimation, then the method provides robustness under various conditions, but lacks theoretical foundation and struggles to distinguish rotational and translational flows
Solution Approach 1:
The geometric solver acts as an intermediary that provides theoretically grounded pose estimates and uncertainty measurements. These geometric estimates serve as a reference that helps the CNN-based system distinguish between rotational and translational flows, reducing the complexity of the deep learning approach while maintaining its robustness.
Solution Approach 2:
The patent implements a feedback mechanism where the geometric solver's uncertainty estimates are used to weight the CNN's predictions. When geometric uncertainty is high (indicating difficult conditions), the CNN's robust estimate is weighted more heavily, and vice versa. This feedback loop allows the system to leverage the CNN's robustness while using geometric theory to guide the estimation process.
3Measurement precision
If uncertainty-based probabilistic fusion is applied, then the accuracy of pose estimation is improved by leveraging strengths of both approaches, but the computational complexity increases
Solution Approach 1:
The patent changes the parameter representation by working with uncertainty (inverse variance) rather than direct pose estimates during fusion. This parameter transformation simplifies the fusion mathematics, allowing the system to compute weighted averages using uncertainty weights without requiring complex probabilistic inference, thus improving accuracy while managing computational complexity.
Data Source
AI summary
A computer-implemented method for fusing geometrical and Convolutional Neural Network (CNN) relative camera pose is provided. The method includes receiving two images having different camera poses. The method further includes inputting the two images into a geometric solver branch to return, as a first solution, an estimated camera pose and an associated pose uncertainty value determined from a Jacobian of a reproduction error function. The method also includes inputting the two images into a CNN branch to return, as a second solution, a predicted camera pose and an associated pose uncertainty value. The method additionally includes fusing, by a processor device, the first solution and the second solution in a probabilistic manner using Bayes' rule to obtain a fused pose.