Pose Regression Training With Decomposed Angle Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques using quaternions for rotation labels in training neural networks for robot localization and mapping are flawed, leading to inefficiencies and inaccuracies in pose regression tasks.
Innovation Solution
Employ a decomposed angle estimator to train and deploy pose regression in neural networks, utilizing a novel representation that considers 4 dimensions instead of the traditional 3 dimensions of Euler angles, enhancing efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If quaternions are used as rotation labels in training methods, then the training process can be completed, but the accuracy and reliability of pose regression deteriorates due to fundamental flaws in quaternion representation
Solution Approach 1:
The patent changes the parameter representation from quaternions to decomposed Euler angles (roll, pitch, yaw) with explicit dimensionality (3D for rotations, 6D for poses). This parameter transformation resolves the fundamental flaws of quaternion representation while maintaining training effectiveness, improving reliability without significantly increasing complexity.
Solution Approach 2:
The patent segments the rotation representation into separate Euler angle components (roll, pitch, yaw) rather than using a unified quaternion approach. This segmentation allows for more intuitive interpretation and better handling of rotation singularities, improving pose regression accuracy while keeping the representation manageable.
2Productivity
If conventional quaternion-based training methods are used, then existing frameworks can be maintained, but efficiency and performance in pose regression tasks deteriorates
Solution Approach 1:
The patent transforms the parameter space from quaternions to decomposed Euler angles, which enables more efficient gradient-based optimization and improves convergence during training. This parameter change enhances both productivity and reliability by eliminating the computational inefficiencies and accuracy issues inherent in quaternion-based methods.
3Ease of operation
If 3D Euler angles are used for rotation representation, then the representation is more intuitive, but it cannot adequately capture the complexity of rotation in 4D pose space
Solution Approach 1:
The patent extends the Euler angle representation from 3D (roll, pitch, yaw) to 6D by adding linear translation components (x, y, z). This dimensional expansion allows the representation to capture both rotational and translational pose information simultaneously, maintaining intuitiveness while significantly improving measurement precision for complete pose regression tasks.
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 one or more images. The method may further include applying pose regression relating to the image capturing device using the real-time view field.