Graph Model Feature Selection for Bundle Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Bundle adjustment processes in image processing face inefficiencies due to unbalanced matched features among images, leading to suboptimal solutions and increased computational expense.
Innovation Solution
A computer-implemented method constructs a graph model based on matched features, assigning capacities and costs to edges, and uses a minimum-cost maximum-flow algorithm to select features as constraints, favoring those viewed by multiple images from different cameras, thereby reducing the number of constraints and balancing connectivity among poses and camera parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an excessive number of matched features are used for bundle adjustment, then the precision of camera poses and intrinsic camera parameters is improved, but the computational expense is significantly increased
Solution Approach 1:
The patent segments the set of matched features into multiple groups based on their distribution across images. A graph model is constructed where nodes represent features and edges represent their occurrence in image pairs. Features are selected for bundle adjustment based on their connectivity in this graph, ensuring balanced representation across multiple images rather than simply using all matched features. This segmentation approach reduces the effective number of constraints while maintaining precision.
Solution Approach 2:
The patent applies local quality by differentiating the role of different matched features based on their image distribution patterns. Features that appear in multiple images from different cameras are prioritized over features appearing in only one or two images. This selective approach ensures that the bundle adjustment process focuses on features that provide the most valuable constraints for improving camera pose precision while avoiding redundant or overly complex computational requirements.
2Productivity
If matched features are unbalanced among images, then the bundle adjustment process can proceed, but suboptimal solutions are obtained
Solution Approach 1:
The patent performs preliminary analysis by constructing a graph model before executing bundle adjustment. In this graph model, nodes represent matched features and edges represent the relationships between features across different images. The graph structure is pre-computed to identify well-distributed features, and this preliminary structuring guides the subsequent feature selection process to ensure balanced representation across all images before the actual optimization begins.
Solution Approach 2:
The patent incorporates feedback mechanisms where the graph model continuously evaluates the distribution of matched features and adjusts the selection process accordingly. By analyzing the connectivity and distribution patterns in the graph, the system identifies which features provide the most balanced constraints across multiple images and prioritizes those for bundle adjustment, ensuring optimal precision while maintaining productivity.
Data Source
AI summary
Systems and methods for adjusting a camera parameter associated with a plurality of two-dimensional images are provided. Matched features can be selected as constraints for bundle adjustment using a graph model constructed based on the matched features. The graph model can have a plurality of nodes including a source node, a sink node, a point node for every matched feature, and a pose pair node associated with each pose pair viewing a matched feature. The graph model can include a plurality of edges interconnecting the plurality of nodes in the graph model. Each of these edges can be assigned a capacity and a cost based on various parameters. A flow can be computed through each of the edges of the graph model using, for instance, a minimum-cost maximum-flow algorithm. Matched features can be selected for bundle adjustment based on the flows of the edges in the graph model.