Partial Point Cloud Pose Estimation With Symmetry-Aware Shape Completion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine vision methods for 6D pose estimation and shape completion face challenges such as high costs for annotated training data, rotational ambiguities with symmetric objects, and inefficiencies in handling symmetry, particularly in industrial robotics applications.
Innovation Solution
A novel encoder-decoder neural network (KNN) trained on CAD models, utilizing a PointNet++ architecture for encoding and a DeepSDF-based decoder to estimate 6D positions and complete shapes from partial point clouds, with separate loss functions for symmetric and asymmetric objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If RGB-based methods are used for 6D pose estimation, then color information is utilized, but additional labeled training data are required which increases costs
Solution Approach 1:
The patent uses synthetic data generated from CAD models as training data, creating virtual copies of real objects in controlled environments. This eliminates the need for expensive annotated real-world images while maintaining training effectiveness through rendered views and synthetic depth information
Solution Approach 2:
The patent introduces depth information as an intermediary representation between RGB images and 6D pose estimation. By training on synthetic depth data and point clouds derived from CAD models, the system bridges the gap between visual input and geometric output without requiring labeled photographs
2Measurement precision
If symmetric objects are processed using conventional 6D pose estimation, then position determination is attempted, but rotational ambiguities occur due to symmetry
Solution Approach 1:
The patent applies asymmetric perturbation to symmetric objects during training by introducing random rotations and transformations. This teaches the network to distinguish between symmetric configurations by learning from artificially created asymmetric variations, enabling accurate pose estimation even for originally symmetric objects
Solution Approach 2:
The patent uses dynamic transformation of training data by applying random rotations, translations, and scaling to CAD models during training. This creates diverse training examples that help the network learn to resolve rotational ambiguities by seeing the same object in many different orientations and configurations
3Adaptability or versatility
If multiple types of training data are used, then comprehensive training is achieved, but training costs and data preparation time increase
Solution Approach 1:
The patent performs preliminary generation of diverse training data from CAD models before actual training begins. By pre-rendering multiple views, depths, and point clouds from standardized CAD files, the system prepares comprehensive training sets automatically without manual annotation, reducing both time and cost
Solution Approach 2:
The patent uses CAD models as a universal source that can generate multiple types of training data (RGB images, depth maps, point clouds, normal maps) from a single model file. This multi-functional approach eliminates the need to collect and annotate separate datasets for each data type, significantly reducing preparation time
Data Source
AI summary
A computer-implemented method for machine vision includes generating, by an artificial neural network, a 6D position estimate of an object and a shape completion estimate of the object based on a partial point cloud. The partial point cloud including a set of 3D points representing the object. The artificial neural network trained using at least one digital geometry model of the object, in particular, at least one CAD model of the object.