Articulated Object Posture Estimation via Iterative Depth Feature Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating the posture of articulated objects in 3D space, such as in video games or virtual worlds, face challenges in accurately determining the position and posture of complex objects like human hands, which affects user immersion and system functionality.
Innovation Solution
A learning method that iteratively learns parameters based on depth features and estimation positions, using labeled training images to compute and refine depth features and position information through multiple iterations, allowing for precise estimation of articulated object postures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current methods are used to estimate posture of articulated objects, then the system can function with simple algorithms, but the accuracy of posture estimation is insufficient
Solution Approach 1:
The patent applies preliminary action by pre-learning parameters from training images before actual posture estimation. The system pre-processes training data to learn depth features and parameters that will be used during runtime, separating the complex learning phase from the simpler estimation phase. This allows high accuracy to be achieved while keeping the runtime algorithm relatively simple.
Solution Approach 2:
The patent implements dynamics by using an iterative refinement process where the estimation is improved through multiple iterations. The algorithm dynamically adjusts the estimation by repeatedly refining depth features and position estimates, allowing the system to adapt and improve accuracy progressively rather than using a static single-step approach.
2Measurement precision
If iterative learning with multiple iterations is performed, then the accuracy of depth feature computation is improved, but the computation time increases
Solution Approach 1:
The patent reduces computation time during runtime by performing the iterative learning process in advance during a training phase. The system pre-computes parameters and depth features from training images, storing them for later use. This shifts the computational burden to an offline phase, making the online estimation faster while maintaining high accuracy.
Solution Approach 2:
The patent applies partial action by performing a limited number of iterations during runtime estimation rather than extensive iterative learning. The system uses pre-learned parameters and performs only necessary refinement iterations, achieving sufficient accuracy without the excessive computation time that would result from performing full iterative learning during runtime.
3Loss of information
If the area for computing depth features is large, then more comprehensive depth information is obtained, but the computational load increases
Solution Approach 1:
The patent applies local quality by focusing computational resources on local regions around the estimation position rather than processing the entire image uniformly. The system generates offset pairs within a localized area centered on the estimation position, concentrating computational effort where it is most needed while reducing overall computational load.
Solution Approach 2:
The patent implements segmentation by dividing the computation into localized regions around different estimation positions. Instead of computing depth features for the entire image at once, the system processes separate local areas around each point of interest, reducing the computational burden while maintaining comprehensive depth information coverage.
Data Source
AI summary
A method of learning a parameter to estimate a posture of an articulated object, and a method of estimating the posture of the articulated object are provided. A parameter used to estimate a posture of an articulated object may be iteratively learned based on a depth feature corresponding to an iteration count, and the posture of the articulated object may be estimated based on the learned parameter.