Neural Network Pose Regression Using Decomposed 6D Angles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques using quaternions for rotation labels in training neural networks for robot localization and mapping (SLAM) are flawed, leading to inefficiencies and inaccuracies in pose regression tasks.
Innovation Solution
Employing a decomposed angle estimator, such as a decomposed 6D angle, to train and deploy pose regressions in neural networks, providing enhanced efficiency and accuracy compared to traditional methods like quaternions, Euler angles, or matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If quaternions are used as rotation labels in training methods, then the neural network can be trained for pose regression, but the training efficiency and accuracy deteriorate due to fundamental flaws in quaternion representation
Solution Approach 1:
The patent changes the parameter representation from quaternions to decomposed 6D angles (three rotation angles and three translation components). This parameter transformation resolves the fundamental flaws of quaternion representation while improving both training efficiency and pose regression accuracy. The decomposed format allows the neural network to learn more effective features without the mathematical constraints of quaternion normalization.
2Ease of manufacture
If quaternions are used for rotation labels, then the training process can proceed, but the rotation representation becomes complex and computationally intensive
Solution Approach 1:
The patent segments the rotation representation into decomposed 6D angles, separating the rotation components from translation components. This segmentation simplifies the overall representation by breaking down the complex quaternion format into more manageable and interpretable angular components, reducing computational intensity while maintaining training feasibility.
3Device complexity
If conventional rotation labels are used, then the neural network structure remains simple, but the pose regression performance suffers from fundamental representation flaws
Solution Approach 1:
The patent applies parameter changes by transforming the rotation label representation from quaternions to decomposed 6D angles. This change improves pose regression accuracy by eliminating the fundamental representation flaws of quaternions while keeping the neural network structure relatively simple and maintainable.
Data Source
AI summary
A mechanism is described for facilitating training and deploying of pose regression in neural networks in autonomous machines. A method, as described herein, includes facilitating capturing, by an image capturing device of a computing device, one or more images of one or more objects, where the one or more images include one or more training images associated with a neural network. The method may further include continuously estimating, in real-time, a present orientation of the computing device, where estimating includes continuously detecting a real-time view field as viewed by the image capturing device and based on the one or more images. The method may further include applying pose regression relating to the image capturing device using the real-time view field.