Stereo Matching Using Atrous CNN for Context Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional stereo matching technologies face challenges in efficiently obtaining depth information from stereo images due to high computational complexity and parameter requirements, especially when dealing with large kernel sizes needed for wide context information extraction.
Innovation Solution
The proposed method employs an Atrous CNN and a stacked 3D encoder-decoder CNN to extract feature maps from stereo images, allowing for effective wide context information retrieval while maintaining spatial resolution and reducing computational complexity, by applying convolutional layers with different Atrous values and performing regression analysis on the cost volume.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If large kernel sizes are used to extract wide context information, then context information extraction is improved, but computational complexity increases
Solution Approach 1:
The patent applies atrous convolution with different dilation rates (1, 2, 4, 8) to extract context information at multiple scales without increasing kernel size. This transforms the approach from spatial expansion (larger kernels) to multi-scale feature extraction through dimensional variation in the feature space, achieving wide context capture with reduced computational burden
Solution Approach 2:
The patent segments the context extraction process into multiple parallel convolutional operations with different atrous rates. Instead of using a single large kernel, the system divides the task into several smaller operations that collectively capture context at various scales, reducing the computational complexity of each individual operation while maintaining overall effectiveness
2Measurement precision
If conventional stereo matching methods are used, then depth information can be obtained, but computational complexity and parameter requirements are high
Solution Approach 1:
The patent replaces conventional mechanical stereo matching algorithms (window-based comparison, cost aggregation) with a deep learning-based CNN system. This substitution leverages learned features and end-to-end optimization to achieve accurate depth estimation with reduced computational complexity, as the network automatically learns optimal feature representations rather than relying on hand-crafted algorithms
Solution Approach 2:
The patent changes the parameter space by using atrous convolution with varying dilation rates instead of fixed kernel sizes. This allows the system to adaptively capture context at multiple scales without increasing the number of parameters significantly, as the same weights are reused across different atrous rates, thereby reducing overall parameter requirements while maintaining depth estimation accuracy
Data Source
AI summary
Disclosed is a stereo matching method for generating a disparity map from a stereo image. The stereo matching method may include obtaining a cost volume by applying a first convolutional neural network (CNN) and a second CNN to a left image and a right image, respectively, wherein the cost volume is determined based on feature maps extracted from the left image and the right image, respectively, performing normalization on the cost volume by applying a third CNN, up-sampling the normalized cost volume, and obtaining a disparity map by applying regression analysis to the up-sampled cost volume.