Multi-Branch 3D Pose Heatmaps for Higher Prediction Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for 3D pose estimation of articulated objects, such as robots and humans, are inefficient in generating semi-perspective decoupled heat maps, leading to suboptimal performance in collision avoidance systems.
Innovation Solution
A deep neural network architecture with multiple branches, each processing features from a 3D scene to generate probabilistic representations, including heat maps, probability density functions, and histograms, to enhance prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single network structure is used to generate all probabilistic representations, then the device complexity is reduced, but the measurement precision and prediction accuracy deteriorate
Solution Approach 1:
The neural network is divided into multiple branches, where each branch is dedicated to generating probabilistic representations for a specific plane (e.g., XY plane, XZ plane, YZ plane). This segmentation allows each branch to specialize in learning features relevant to its specific plane, thereby improving prediction accuracy for each plane while maintaining a manageable overall structure through modular organization.
2Measurement precision
If multiple branches with different weights are used to generate probabilistic representations for different planes, then the prediction accuracy improves, but the device complexity increases
Solution Approach 1:
Each branch in the multi-branch network is assigned different local weights and parameters optimized for its specific plane. For example, the branch for the XY plane has weights tailored to capture features relevant to horizontal positioning, while the branch for the XZ plane has different weights for vertical positioning. This local quality approach enables each branch to achieve high prediction accuracy for its designated plane without requiring the entire network to be optimized for all planes simultaneously, thus managing complexity through specialization.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure provides a computer-implemented method for generating probabilistic representations. The method includes a step of processing, using branches of a deep neural network, features derived from a data representation of a three-dimensional "3D" scene to generate different groups of probabilistic representations. Each of the branches comprises one or more layers that include a different set of weights for generating one of the different groups from the features. Each of the groups is associated with a different plane of a 3D space and comprises one or more probabilistic representations. A probabilistic representation in one of the groups indicates probability that a projection of a target in the 3D scene to the plane associated with the group lies in one or more locations of the plane.