Robust Feature Matching Using Invariant Mutual Relations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for matching objects in images, particularly in three-dimensional real-world scenes, face limitations due to the need for descriptor vectors that are invariant to specific transformations, leading to reduced uniqueness and descriptive power, and relational matching techniques have struggled to reliably handle such scenes.
Innovation Solution
The system employs robust feature matching using mutual relations between features, where sets of invariant relations are defined and used to determine object positions, applicable to various dimensional projections from sensors like LIDAR and cameras, allowing for accurate transformation estimation between image captures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If descriptor vectors are made more generic to be invariant to all transformations, then adaptability improves, but object uniqueness and descriptive power deteriorate
Solution Approach 1:
The patent segments the scene into multiple objects and represents each object with its own local coordinate system and feature set. Instead of using a single generic descriptor for the entire scene, the system divides the representation into object-specific segments, each maintaining its own invariant properties. This allows the system to adapt to different transformations while preserving the unique characteristics of each object through its segmented representation.
2Loss of information
If descriptor vectors are tuned to a specific situation, then object uniqueness improves, but adaptability to other situations deteriorates
Solution Approach 1:
The patent creates a universal object representation framework that can handle multiple situations and transformations. Each object is represented with a coordinate system and feature set that serves multiple functions: it maintains uniqueness for the specific object while also being adaptable to various transformations and viewing conditions. The mutual relation matching mechanism provides a universal approach that works across different scenes and configurations.
3Measurement precision
If relational matching techniques are used instead of description vectors, then object relationship accuracy improves, but reliability in handling three-dimensional real world scenes deteriorates
Solution Approach 1:
The patent introduces mutual relations between objects as an intermediary mechanism that connects local object features to global scene understanding. Instead of directly matching objects across 3D scenes without intermediate structure, the system uses defined mutual relations (spatial, temporal, or semantic) as mediators. This intermediary layer enables accurate relationship matching while maintaining reliability in 3D real-world scenes by providing a structured framework for interpreting object interactions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the present invention provide improved systems and methods for matching scenes. In one embodiment, a processor for implementing robust feature matching between images comprises: a first process for extracting a first feature set from a first image projection and extracting a second feature set from a second image projection; a memory for storing the first feature set and the second feature set; and a second process for feature matching using invariant mutual relations between features of the first feature set and the second feature set; wherein the second feature set is selected from the second image projection based on the identification of similar descriptive subsets between the second image projection and the first image projection.