Stereo Disparity Confidence Measure for False Disparity Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stereo disparity estimation algorithms face challenges with inaccurate results and high computational costs, and the fusion of multiple disparity maps is unreliable due to propagated false disparities, leading to suboptimal disparity maps.
Innovation Solution
A new confidence measure is introduced that evaluates disparity candidates by accumulating contribution values based on dissimilarity and distance from the global minimum, effectively filtering out false disparities by decreasing confidence and increasing uncertainty with increasing local minima and disparity distance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional confidence measures are used to evaluate disparity candidates, then the disparity map generation can proceed, but false disparities are propagated leading to unreliable disparity maps
Solution Approach 1:
The patent implements a feedback mechanism where the confidence measure evaluates the cost curve characteristics and feeds back uncertainty information to filter false disparities. The confidence measure analyzes the shape of cost curves (number of local minima, depth of global minimum) and uses this feedback to identify and exclude unreliable disparity candidates, thereby improving the overall reliability of the disparity map.
Solution Approach 2:
The confidence measure acts as an intermediary between the disparity estimation algorithm and the final disparity map generation. It introduces an additional evaluation layer that assesses the quality of disparity candidates by analyzing cost curve characteristics, mediating which disparities are accepted or rejected before final map construction, thus preventing false disparities from propagating.
2Measurement precision
If deep-learning methods are used to improve disparity estimation accuracy, then measurement precision increases, but computational costs increase significantly
Solution Approach 1:
The patent employs a computationally efficient confidence measure that does not require heavy deep-learning models. Instead of using expensive trained networks, it uses a lightweight analytical approach based on cost curve characteristics (counting local minima, measuring their depth), which is much cheaper computationally while still effectively identifying false disparities.
Solution Approach 2:
The patent replaces the complex mechanical/deep-learning-based disparity evaluation system with a simpler analytical system. Instead of using neural networks to evaluate disparity quality, it substitutes this with a mathematical analysis of cost curve properties, reducing computational overhead while maintaining effectiveness.
3Measurement precision
If multiple local minima are present in the cost curve, then disparity localization becomes ambiguous, but the confidence measure should still be able to identify the global minimum
Solution Approach 1:
The patent applies a criterion that requires the global minimum to be significantly deeper than other local minima (excessive depth requirement). By setting a threshold on how much deeper the global minimum must be compared to other minima, it filters out cases where multiple minima have similar depths, thereby reducing ambiguity and ensuring only clear, unambiguous disparity localizations are accepted.
Data Source
Figure 1
Figure 2a~2b
Figure 2c~2d
AI summary
A more effective confidence/uncertainty measure determination for disparity measurements is achieved by performing the determination on an evaluation of a set of disparity candidates for a predetermined position of a first picture at which the measurement of the disparity relative to the second picture is to be performed, and if this evaluation involves an accumulation of a contribution value for each of this set of disparity candidates, which contribution values depends on the respective disparity candidate and a dissimilarity to the second picture which is associated with the respective disparity candidate according to a function which has a first monotonicity with a dissimilarity associated with the respective disparity candidate, and a second monotonicity, opposite to the first monotonicity, with an absolute difference between the respective disparity candidate and a predetermined disparity having a minimum dissimilarity associated therewith among dissimilarities associated with the set of disparity candidates. By this means, the confidence/uncertainty measure tends to decrease the confidence, and increase the uncertainty, with increasing number of local minima in the spatial distribution of disparity candidates. Further, the larger the disparity distance of any local minimum to the global minimum in terms of disparity is, the lower the confidence and the higher the uncertainty, respectively, tends to be. Further, the larger the distance of any local minimum to the global minimum in terms of dissimilarity is, the lower is its influence of the confidence/uncertainty measure determination.