Stereo Matching via Unary and Pairwise Confidence Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Stereo matching techniques face challenges in accurately determining binocular disparity due to similarities in pixel colors and sizes, leading to errors in minimizing cost functions and incorrect disparity information.
Innovation Solution
A stereo matching method and apparatus that determine pairwise and unary confidence measures based on similarities and discontinuities between pixels in left and right images, using confidence learners to extract and learn relationships, thereby refining the cost function for accurate stereo matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stereo matching is performed using traditional cost function minimization, then binocular disparity information can be obtained, but errors occur due to pixel similarity and incorrect matching
Solution Approach 1:
The patent introduces confidence feedback mechanisms where unary confidence (pixel-level matching reliability) and pairwise confidence (local neighborhood consistency) are computed and fed back into the cost function. This feedback loop allows the system to iteratively refine disparity estimates by weighting matches based on their confidence scores, thereby improving both measurement precision and reliability simultaneously
Solution Approach 2:
The patent transforms the traditional binary matching problem into a probabilistic framework by introducing confidence parameters (unary and pairwise confidence values). These parameters continuously adjust the cost function weights based on local image characteristics, allowing the system to adaptively handle varying pixel similarities and occlusions, thus resolving the contradiction between accuracy and reliability
2Productivity
If cost function minimization is used for stereo matching, then binocular disparity can be determined, but errors occur in minimizing the cost function leading to incorrect disparity information
Solution Approach 1:
The patent performs preliminary confidence assessment before final disparity determination. Unary confidence is computed at the pixel level and pairwise confidence is calculated for local neighborhoods beforehand. These pre-computed confidence measures guide the subsequent cost function minimization process, ensuring that efficiency is maintained while accuracy is improved through informed weighting
3Device complexity
If traditional stereo matching is performed without confidence learning, then the process is simpler, but errors occur due to pixel similarities in color and size
Solution Approach 1:
The patent segments the matching reliability assessment into two distinct components: unary confidence (pixel-level characteristics) and pairwise confidence (local neighborhood relationships). This segmentation allows the system to independently evaluate different aspects of matching quality without significantly increasing overall complexity, as each component can be computed using dedicated algorithms
Data Source
AI summary
A stereo matching apparatus and method through learning a unary confidence and a pairwise confidence are provided. The stereo matching method may include learning a pairwise confidence representing a relationship between a current pixel and a neighboring pixel, determining a cost function of stereo matching based on the pairwise confidence, and performing stereo matching between a left image and a right image at a minimum cost using the cost function.