Non-Uniform Resampling for 3D Point Cloud Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for automated semantic segmentation of non-Euclidean 3D data sets, such as point clouds from intra-oral scanners, face challenges due to irregularity, permutation invariance, and variable point numbers, leading to difficulties in accurately classifying fine high-resolution details like tooth and gingiva borders.
Innovation Solution
A non-uniform resampling algorithm generates locally-dense and globally-sparse subsets of points, which are then fed into a deep neural network for semantic segmentation, allowing the network to learn both local fine details and global shape structures, and an adversarial training setup with a discriminator network improves label accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If uniform resampling is used to process point clouds, then processing complexity is reduced, but local fine details (tooth and gingiva borders) are lost
Solution Approach 1:
The patent applies local quality by implementing a non-uniform resampling strategy where points are densely sampled in regions containing fine geometric details (such as tooth and gingiva borders) and sparsely sampled in regions with coarser geometry. This allows the system to preserve measurement precision where needed while maintaining manageable processing complexity elsewhere.
2Measurement precision
If all points in the point cloud are processed, then segmentation accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies segmentation by dividing the point cloud into multiple subsets based on geometric criteria, such as regions containing fine details versus regions with coarser geometry. This allows the deep learning model to process different subsets with appropriate sampling densities, improving segmentation accuracy while reducing overall processing time compared to processing all points uniformly.
3Measurement precision
If non-uniform resampling is used to preserve fine details, then segmentation accuracy is improved, but data irregularity increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the sampling density parameter based on local geometric properties of the point cloud. Regions with high curvature or fine geometric features receive higher sampling density, while regions with smoother surfaces receive lower density. This adaptive parameter change preserves segmentation accuracy while managing data irregularity through a systematic approach.
4Measurement precision
If high-resolution point clouds are processed, then fine detail preservation is improved, but computational resources required increase
Solution Approach 1:
The patent applies local quality by concentrating computational resources on processing points in regions containing fine geometric details, while using fewer resources for regions with coarser geometry. This is achieved through non-uniform resampling that adaptively allocates more processed points to areas requiring higher precision, thereby preserving fine detail while optimizing computational resource usage.
Data Source
AI summary
A computer-implemented method for semantic segmentation of a point cloud comprises receiving a cloud having points representing a vector of an object, preferably part of a dento-maxillofacial structure having a dentition; determining subset(s) including a first number of points arranged around a selected point of the cloud and a second number of points arranged at spatial distances larger than a predetermined spatial distance of the first number of points, the first number of points representing fine feature(s) of the object around the selected point and the second number of points representing object global feature(s); providing each subset of points to a deep neural network, DNN, the DNN being trained to semantically segment points of each subset according to classes associated with the object; and, for each subset point, receiving a DNN output multi-element vector, wherein each element represents a probability that the point belongs to class(es) of the object.