3D Object Segmentation via Iterative Primitive Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for segmenting 3D modeled objects, particularly those representing mechanical assemblies, face challenges in achieving accurate and robust segmentation with reduced computation time, as they often rely on supervised learning that limits application to unlabeled and unseen object categories, and unsupervised methods struggle with semantically relevant shape segmentations.
Innovation Solution
A computer-implemented method involving an abstraction neural network that outputs an initial set of primitives, followed by an iterative merging process to reduce the number of primitives, using a merging neural network to combine subsets into single primitives, thereby achieving a more accurate and semantically relevant segmentation of 3D modeled objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised deep learning methods are used to segment 3D shapes, then segmentation accuracy is improved, but the method cannot be applied to unlabeled and unseen object categories
Solution Approach 1:
The patent inverts the conventional supervised learning approach by using unsupervised learning methods. Instead of requiring labeled data to train the network, the system learns from raw 3D shape data without labels, enabling segmentation of unseen object categories while maintaining reasonable segmentation accuracy through iterative refinement processes.
Solution Approach 2:
The patent changes the learning paradigm parameter from supervised to unsupervised learning. This fundamental parameter change allows the system to process unlabeled 3D data by learning intrinsic structures and patterns directly from the data distribution, thereby achieving versatility across different object categories without requiring category-specific training labels.
2Adaptability or versatility
If unsupervised learning methods are used to segment 3D objects, then adaptability to unseen categories is improved, but segmentation precision and semantic relevance deteriorate
Solution Approach 1:
The patent applies preliminary action by first generating an initial set of primitives through unsupervised learning, then iteratively refining these primitives through merging operations. This preliminary segmentation provides a starting point that can be progressively improved, allowing the system to achieve better precision while maintaining adaptability to unseen categories.
Solution Approach 2:
The patent introduces dynamics through the iterative refinement process. The segmentation is not static but evolves through multiple iterations where primitives are merged and refined based on geometric consistency and semantic relevance criteria. This dynamic refinement allows the system to improve segmentation precision progressively while maintaining the flexibility to handle diverse object categories.
3Productivity
If a fixed number of parts is output by the segmentation method, then computation time is reduced, but semantic relevance of shape segmentation is lost
Solution Approach 1:
The patent makes the number of segmented parts dynamic rather than fixed. The iterative refinement process allows the segmentation to evolve from an initial fixed-number primitive set to a final segmentation with a variable number of parts that better reflects the semantic structure of the input shape. This dynamic adjustment enables the system to achieve semantic relevance while managing computation time through efficient iteration.
Solution Approach 2:
The patent applies partial action by performing a limited number of iterative refinement steps rather than exhaustively optimizing all possible segmentations. This partial refinement approach achieves a reasonable balance between computation time and semantic relevance, producing segmentations that are sufficiently accurate without requiring excessive computational resources.
4Reliability
If iterative merging of primitives is performed to reduce primitive count, then segmentation robustness is improved, but computation time increases
Solution Approach 1:
The patent applies preliminary action by generating an initial set of primitives that already capture the major structural features of the 3D shape. This preliminary segmentation reduces the number of iterations needed for refinement, thereby improving robustness while limiting the increase in computation time. The initial primitives serve as a strong foundation that requires fewer refinement steps to achieve reliable segmentation.
Solution Approach 2:
The patent performs a limited number of merging iterations rather than exhaustively merging all possible primitive combinations. This partial iteration approach achieves sufficient robustness for practical applications while avoiding the excessive computation time that would result from complete iterative optimization. The system stops iterations when diminishing returns are detected, balancing robustness and efficiency.
Data Source
AI summary
The disclosure relates to a computer-implemented method comprising inputting a representation of a 3D modeled object to an abstraction neural network which outputs a first set of a first number of first primitives fitting the 3D modeled object; and determining, from the first set, one or more second sets each of a respective second number of respective second primitives. The second number is lower than the first number. The determining includes initializing a third set of third primitives as the first set and performing one or more iterations, each comprising to merging one or more subsets of third primitives together each into one respective single fourth primitive, to thereby obtain a fourth set of fourth primitives. Each iteration further comprises setting the third set of a next iteration as the fourth set of a current iteration and setting the one or more second sets as one or more obtained fourth sets.