Cascaded Feature Extraction for Stereo Vision Disparity Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current disparity prediction methods in binocular stereo vision face challenges in accuracy, particularly in areas with repetitive patterns, missing textures, and complex edges, and are computationally inefficient due to the need for repeated neural network runs during refinement.
Innovation Solution
A method involving cascaded feature extraction layers to iteratively refine disparity maps using residual matching costs, reducing error and improving accuracy without significantly increasing computational load, by constructing an initial disparity map and refining it layer-by-layer with attention mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional disparity prediction methods are used, then the computational process is simpler, but the accuracy of disparity maps deteriorates in areas with repetitive patterns, missing textures, and complex edges
Solution Approach 1:
The patent divides the disparity prediction process into multiple cascaded stages (e.g., coarse-to-fine pyramid structure), where each stage processes the disparity map at a different resolution level. This segmentation allows the system to handle complex regions with higher precision at finer levels while maintaining computational efficiency at coarser levels, directly resolving the contradiction between accuracy and complexity.
Solution Approach 2:
The patent introduces an additional dimensional aspect by processing disparity maps across multiple scales and resolutions rather than a single level. This multi-scale approach enables the system to capture both global context and local details, improving accuracy in challenging regions without proportionally increasing computational complexity at any single level.
2Measurement precision
If multiple neural network runs are performed for refinement, then the accuracy of disparity maps is improved, but the computational time and efficiency deteriorate
Solution Approach 1:
The patent performs preliminary disparity estimation at coarser resolution levels before refining at finer levels. This preliminary action provides a good initial estimate that guides subsequent refinement steps, reducing the computational burden of multiple full-resolution neural network runs while maintaining high accuracy through progressive refinement.
Solution Approach 2:
The patent implements periodic refinement at different pyramid levels rather than continuous full-resolution processing. The system periodically upsamples and refines the disparity map at each level, creating a rhythm of coarse-to-fine processing that maintains accuracy while improving computational efficiency by avoiding constant full-resolution operations.
3Speed
If single-scale disparity estimation is used, then the processing speed is faster, but the accuracy deteriorates in regions with varying depths and complex structures
Solution Approach 1:
The patent segments the image processing into multiple scales using a pyramid structure, where each level processes different spatial frequencies and depth ranges. This segmentation enables simultaneous processing of both global scene structure at coarse levels and local detailed structures at fine levels, achieving high accuracy without sacrificing overall processing speed.
Solution Approach 2:
The patent adds the scale dimension to the processing pipeline, transforming single-scale estimation into multi-scale analysis. This dimensional expansion allows the system to capture depth variations at different granularities, improving accuracy in complex regions while maintaining processing efficiency through hierarchical organization.
Data Source
AI summary
Embodiments of the disclosure provide a disparity map acquisition method and apparatus, a device, a control system and a storage medium. The method includes: respectively performing feature extraction on left-view images and right-view images of a captured object layer by layer through M cascaded feature extraction layers, to obtain a left-view feature map set and a right-view feature map set of each layer, M being a positive integer greater than or equal to 2; constructing an initial disparity map based on the left-view feature map set and the right-view feature map set extracted by an Mth feature extraction layer; and iteratively refining, starting from an (M−1)th layer, the disparity map through the left-view feature map set and the right-view feature map set extracted by each feature extraction layer in sequence until a final disparity map is obtained based on an iteratively refined disparity map of a first layer.