AR Pose Estimation Using Local Keypoint Neural Networks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing 3D interaction technologies for augmented reality (AR) have limitations in pose estimation accuracy due to reliance on global image information from a single image, which is insufficient for high-accuracy applications.

Innovation Solution

A pose estimation method that utilizes local information, such as keypoint information, through a neural network with shared parameters across tasks for estimating pose, keypoint, and class information, employing convolution and fully-connected layers in parallel or cascade modes to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If pose estimation is performed using global information from a single image, then the method is simple to implement, but the pose estimation accuracy is insufficient

Engineering Contradiction:

Improvepose estimation accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the image processing task by introducing keypoint detection to identify specific local features (corners, edges, distinctive points) rather than processing the entire image globally. This segmentation of attention to critical regions enables more accurate pose estimation while maintaining computational efficiency through focused processing of key areas.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by using keypoint information to capture local image characteristics that are critical for accurate pose estimation. By focusing on specific local features (keypoints) rather than uniform global processing, the system achieves higher measurement precision in pose estimation while managing complexity through selective local analysis.

Inventive Principle:

Principle #3Local quality

2Measurement precision

If multiple tasks (pose estimation, keypoint detection, class recognition) are performed separately, then each task can be optimized independently, but the overall system complexity increases

Engineering Contradiction:

Improveestimation accuracyVSAvoidnetwork complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple tasks (pose estimation, keypoint detection, and class recognition) into a single unified neural network architecture. This combination allows the network to share computational resources and learn correlated features across tasks simultaneously, improving overall estimation accuracy while reducing redundancy and managing system complexity through integrated processing.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The neural network is designed with multi-functionality to perform multiple tasks (pose estimation, keypoint detection, class recognition) within a single unified framework. This universal approach allows the network to learn shared representations that benefit all tasks, achieving high estimation accuracy across multiple objectives while avoiding the complexity of separate specialized systems.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS10937189B2Pose estimation method, method of displaying virtual object using estimated pose, and apparatuses performing the same

Publication Date: 2021.03.02 SAMSUNG ELECTRONICS CO LTD

AI summary

Disclosed are a pose estimation methods and apparatuses of displaying a virtual object using an estimated pose. The pose estimation method includes receiving an input image and estimating pose information of an object from the input image based on local information of the object.