3D Model Pose Estimation via Depth Difference Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Object detection and pose estimation techniques face challenges in accurately estimating the pose of objects that overlap or are obscured by background clutter, leading to significant missing data and increased processing time in automatic bin picking scenarios.
Innovation Solution
A method involving a non-transitory computer readable medium that trains an ensemble of decision trees by generating multiple views of a 3D model, calculating depth differences, and associating perturbations to refine object pose estimation, using a pyramid of ensemble trees for accurate pose refinement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If geometrical template matching using local invariant features or point pair features is used for pose estimation, then the system can identify objects in cluttered environments, but the processing time increases significantly and accuracy decreases when objects overlap
Solution Approach 1:
The system pre-generates multiple views of the 3D model and pre-calculates depth maps and reference 2D locations before actual pose estimation is needed. This preliminary preparation creates a library of expected object appearances from different angles, allowing the system to quickly match against pre-computed data rather than performing complex calculations in real-time during object detection
Solution Approach 2:
The patent segments the pose estimation problem into multiple independent components: generating multiple views of the object, calculating depth maps for each view, identifying reference 2D locations in overlapping areas, transforming these locations into 3D-point sets, and calculating depth differences. By dividing the complex pose estimation task into these smaller, pre-computable segments, the system reduces real-time processing requirements while maintaining accuracy
2Measurement precision
If multiple views of 3D model are generated and depth differences are calculated to refine pose estimation, then the accuracy of pose estimation is improved, but the complexity of the learning function increases
Solution Approach 1:
The system creates multiple copies or views of the 3D model from different angles and positions. Instead of developing a single complex learning function to handle all possible object orientations and overlaps, the patent generates multiple simplified view-specific models. Each view provides a different perspective on the object, and the system learns from multiple simpler copies rather than one overly complex function, reducing overall learning complexity while improving precision through multi-view consensus
Data Source
AI summary
Training a pose detection algorithm and deriving an object pose using a trained pose detection algorithm. Specifying a first view of a 3D model with respect to a virtual camera, generating a depth map by projecting the 3D model based on the first view, generating a second view by perturbing the first view, selecting reference 2D locations in an overlapping area of the 3D model between the first view and the second view. Transforming the reference 2D locations into a 3D point set corresponding to the second view. For each reference 2D location, calculating a depth difference between the first view and the second view at the reference 2D location. Associating the perturbation between the second view and the first view with the calculated depth differences, and training an algorithm model to identify a pose of an object corresponding to the 3D model.