Multi-Camera Image Augmentation for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models trained on images captured from limited viewpoints, such as low-angle views, face accuracy and stability issues due to the lack of comprehensive training data sets, particularly in devices like robot vacuum cleaners that require accurate object recognition for tasks.
Innovation Solution
A method to train a neural network model by augmenting images captured by multiple cameras, utilizing conversion relationships between different camera coordinate systems to generate training data, enabling accurate object recognition across varied viewpoints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If neural network models are trained on images captured from limited viewpoints, then device complexity is reduced, but object recognition accuracy deteriorates
Solution Approach 1:
The patent creates virtual copies of training images by applying coordinate transformations and viewpoint conversions. Instead of physically capturing images from multiple viewpoints, the system generates synthetic images representing different camera perspectives through mathematical transformations, thereby improving training data diversity without adding physical cameras.
Solution Approach 2:
The patent transforms 2D images into multiple virtual viewpoints by introducing a third dimension of perspective through coordinate system conversions. The system converts images from one camera coordinate system to another, effectively adding viewpoint diversity without physical movement or additional hardware.
2Measurement precision
If multiple cameras are used to capture images from different viewpoints, then object recognition accuracy is improved, but device complexity increases
Solution Approach 1:
The patent creates virtual copies of training images by applying coordinate transformations and viewpoint conversions. Instead of physically capturing images from multiple viewpoints, the system generates synthetic images representing different camera perspectives through mathematical transformations, thereby improving training data diversity without adding physical cameras.
Solution Approach 2:
The patent replaces the mechanical system of multiple physical cameras with a computational approach. Instead of using multiple cameras to capture images from different viewpoints, the system uses coordinate transformations and viewpoint conversion algorithms to generate virtual images from a single camera's captured data.
3Loss of time
If training data is collected from limited viewpoints, then data collection time is reduced, but neural network model performance deteriorates
Solution Approach 1:
The patent performs preliminary coordinate transformations and viewpoint conversions during the training data preparation phase. By pre-processing images to generate multiple virtual viewpoints from single captured images, the system eliminates the need for time-consuming multi-viewpoint data collection while ensuring comprehensive training data coverage.
Solution Approach 2:
The patent creates virtual copies of training images by applying coordinate transformations and viewpoint conversions. Instead of physically capturing images from multiple viewpoints, the system generates synthetic images representing different camera perspectives through mathematical transformations, thereby improving training data diversity without adding physical cameras.
Data Source
AI summary
Provided is a computer-implemented method of training a neural network model by augmenting images representing objects. The method includes: obtaining a first object recognition result predicted by a first neural network model using, as an input, a first image captured by a first camera capturing, from a first viewpoint, a space including at least one object; converting the obtained first object recognition result, based on a conversion relationship between a first camera coordinate system corresponding to the first camera and a second camera coordinate system corresponding to a second camera capturing, from a second viewpoint, the space; generating, based on the first object recognition result converted with respect to the second viewpoint, training data by performing labeling on a second image that corresponds to the first image, the second image being captured by the second camera; and training a second neural network model by using the generated training data.