Subpixel Disparity Estimation Using Neural Networks and Synthetic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing subpixel disparity calculation methods, such as parabola fitting and neural networks relying on parabola fitting, suffer from systematic bias and noise-induced randomness, and are computationally expensive, especially in real-time applications like SLAM, lacking robustness and accuracy due to reliance on biased distance metrics and insufficient training data.
Innovation Solution
A computer-implemented method using a neural network trained in a supervised manner with both realistic and synthetically generated image data, performing end-to-end prediction directly on image patches, reducing randomness and improving accuracy by simulating camera movements through non-linear operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If parabola fitting is used for subpixel disparity calculation, then the calculation process is simple, but systematic bias and pixel locking occur resulting in reduced accuracy
Solution Approach 1:
The patent replaces the traditional parabola fitting mathematical model with a neural network-based learning system. Instead of using deterministic mathematical formulas that inherently produce bias, the system uses trained neural networks to learn subpixel disparities from image data, substituting mechanical calculation with adaptive learning to eliminate systematic errors while maintaining computational efficiency.
Solution Approach 2:
The patent changes the fundamental parameters of the calculation system by using multiple distance metrics (SAD, SSD, NMSE) as input features to the neural network, rather than relying on a single metric. This multi-parameter approach allows the network to learn more robust disparity estimates and overcome the limitations of individual metrics that contribute to bias in traditional methods.
2Use of energy by moving object
If traditional distance metrics (SAD, SSD) are used for subpixel disparity calculation, then the computation cost is low, but randomness and noise are introduced reducing reliability
Solution Approach 1:
The patent merges multiple distance metrics (SAD, SSD, NMSE) and their combinations as input features to the neural network. By combining these metrics rather than using them separately, the system leverages their complementary strengths and reduces the randomness and noise inherent in individual metrics, achieving more reliable disparity estimates without proportionally increasing computation cost since the neural network processes them efficiently.
Solution Approach 2:
The neural network acts as an intermediary between the distance metrics and the final disparity output. Instead of directly using distance metrics that introduce noise, the network learns optimal transformations and combinations, filtering out randomness while preserving reliable signal, thus mediating between low-cost computation and high reliability requirements.
3Measurement precision
If neural networks are used for subpixel disparity calculation, then accuracy can be improved, but computation cost increases significantly
Solution Approach 1:
The patent uses a relatively simple neural network architecture with a limited number of layers and neurons, applying partial action rather than excessive complexity. The network is designed to be just sufficient for the task - using basic fully connected layers without overly complex architectures - thereby achieving the necessary accuracy improvement over traditional methods while controlling computation cost for practical deployment.
Solution Approach 2:
The patent optimizes network parameters including the number of layers, neurons per layer, learning rate, and batch size to achieve the best balance between accuracy and computation cost. By carefully tuning these parameters, the system achieves high accuracy without requiring excessive computational resources, making the approach feasible for real-world applications.
4Loss of time
If neural networks are trained with insufficient or poorly prepared training data, then training is faster, but the model lacks robustness and generalization ability
Solution Approach 1:
The patent performs preliminary actions in the training process by carefully preparing and curating high-quality training data before actual training begins. This includes selecting appropriate image pairs, ensuring proper alignment, and preparing ground truth labels in advance, which enables efficient training with smaller datasets while maintaining model robustness and generalization ability.
Data Source
AI summary
In a method for subpixel disparity calculation, image data for various images each representing a field of view of an input device is received by a processor, and the image data is applied to a machine learning model. The machine learning module uses the image data to compute an output representing calculated subpixel disparity between the various images. In an example of the method, the machine learning model is a neural network that produces accurate and reliable subpixel disparity estimation in real-time using synthetically generated data.