Real-Time Depth Extraction Using 1D Stereo Disparity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for extracting depth information from stereo images are unstable, require high computing power, and are unsuitable for real-time processing, making them ineffective for graphics-related applications such as scene rendering and object detection.
Innovation Solution
A computer-implemented method that determines one-dimensional disparity between rectified images to extract depth information, using a multi-resolutional data structure and local approximation formulas to simplify depth extraction and enable real-time analysis, allowing for the construction of three-dimensional images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional techniques are used to extract depth information from stereo images, then depth information can be obtained, but the results are non-stable and require high computing power
Solution Approach 1:
The patent divides the depth extraction problem into two distinct phases: a training phase where a neural network is trained offline to learn correspondence relationships, and an inference phase where the trained network performs rapid depth extraction in real-time. This segmentation allows the computationally intensive training to be done once, while real-time processing uses the pre-trained model for stable and efficient depth extraction without requiring high computing power during operation.
Solution Approach 2:
The patent performs preliminary training of the neural network model in advance before actual depth extraction is needed. The training phase pre-computes and stores the correspondence relationships between stereo image pixels, creating a trained model that can be directly applied during real-time processing. This preliminary action eliminates the need for complex computations during real-time depth extraction, providing both stability and efficiency.
2Productivity
If conventional techniques are used to extract depth information from stereo images, then depth information can be obtained, but the processing is unsuitable for real-time applications
Solution Approach 1:
The patent segments the processing into offline training and online inference stages. The computationally heavy neural network training is performed offline in advance, creating a pre-trained model. During real-time operation, only the lightweight inference process is needed, which rapidly processes stereo images to extract depth information. This enables real-time processing capability without requiring sustained high computing power during actual depth extraction.
Solution Approach 2:
The neural network model is trained in advance during an offline preparation phase, storing learned correspondence patterns in a pre-trained model. When real-time depth extraction is required, the system simply applies this pre-computed model to new stereo image pairs, enabling fast processing. This preliminary computation of the training phase allows the system to achieve real-time performance with minimal computing power during operation.
3Measurement precision
If pixel-to-pixel correspondence is established between stereo images, then depth information can be extracted, but the process demands high computing power
Solution Approach 1:
The patent replaces traditional mechanical/computational correspondence matching algorithms with a neural network-based approach. Instead of using conventional image matching techniques that require extensive pixel-to-pixel comparison and computation, the system uses a trained neural network that has learned correspondence relationships during offline training. This substitution maintains measurement precision for depth extraction while dramatically reducing the computing power required during real-time operation.
Data Source
AI summary
Systems and methods may provide for determining a one-dimensional (1D) disparity between a plurality of rectified images, and extracting depth information from the plurality of rectified images based at least in part on the 1D disparity. In one example, the 1D disparity is in the horizontal direction and the images are rectified with respect to one another in the vertical direction.