Disparity Map Refinement With Guide Image Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing disparity and depth estimation methods, particularly those using deep learning networks, often produce inaccurate and sparse results with artifacts and feature matching errors, failing to refine disparity and depth maps effectively.
Innovation Solution
A system that refines disparity and depth maps by processing a reference image and a guide image to generate features, modifying these features using a guided disparity-modulation model, and generating a modified data map with increased accuracy and density, optionally using confidence-based fusion to combine the refined and initial maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning networks are used for disparity and depth estimation, then automation is improved, but measurement precision deteriorates due to artifacts and feature matching errors
Solution Approach 1:
The patent introduces a guide image as an intermediary element that mediates between the reference image and the disparity map generation process. The guide image provides additional structural information that helps correct feature matching errors and reduces artifacts in the generated disparity and depth maps, thereby improving measurement precision while maintaining automation.
2Productivity
If existing deep learning methods are used, then productivity is improved, but manufacturing precision deteriorates due to sparse and inaccurate results
Solution Approach 1:
The patent merges multiple input images (reference image and guide image) to generate the disparity and depth maps. This combination of multiple information sources enables more accurate and dense results compared to using single-image deep learning methods, improving manufacturing precision while maintaining computational efficiency.
3Device complexity
If simple feature matching is used, then device complexity is reduced, but measurement precision deteriorates due to artifacts and errors
Solution Approach 1:
The patent adds another dimension to the feature matching process by incorporating the guide image as an additional input channel. This extra dimensional information helps disambiguate difficult matching cases and reduces artifacts without significantly increasing device complexity, as the additional processing follows the same neural network architecture.
Data Source
AI summary
Systems and techniques are described herein for modifying a map. For instance, a method for modifying a map is provided. The method may include processing a first data map and a reference image to generate first features, the first data map including a first number of data values; processing a guide image to generate second features; modifying the first features based on the second features to generate modified features; modifying the first data map based on the modified features to generate a modified first data map; and generating a second data map based on the modified first data map, the second data map including a second number of data values that is greater than the first number of data values.