Point Cloud Segmentation via Shared Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current point cloud segmentation methods relying on feature-based similarity matrices are inefficient, especially when dealing with large point clouds, leading to low efficiency and accuracy in segmentation processing.
Innovation Solution
A method that encodes point clouds to obtain shared features, decodes them using different branches to get semantic and instance features, and fuses these features to achieve instance segmentation with rich semantic awareness, eliminating the need for feature-based similarity matrices and enhancing segmentation efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature-based similarity matrix is used for point cloud segmentation, then segmentation can be performed, but processing time increases significantly and efficiency decreases
Solution Approach 1:
The patent replaces the traditional mechanical computation of feature-based similarity matrices with a neural network-based feature extraction and fusion system. The neural network automatically learns hierarchical features and performs segmentation through learned transformations, eliminating the need for explicit similarity matrix computations and significantly reducing processing time while maintaining accuracy.
Solution Approach 2:
The patent changes the parameter representation from traditional hand-crafted features to learned neural network features. By transforming the input point cloud through multiple neural network layers with different decoding branches, the system learns optimized feature representations that enable efficient segmentation without relying on computationally intensive similarity matrices.
2Productivity
If feature-based similarity matrix is used for point cloud segmentation, then segmentation can be performed, but segmentation efficiency becomes low especially for large point clouds
Solution Approach 1:
The patent segments the feature extraction process into multiple independent decoding branches that process different aspects of the point cloud simultaneously. Each branch handles specific feature types (geometric, semantic, instance) in parallel, dividing the computational complexity into manageable components that can be processed efficiently even for large point clouds.
Solution Approach 2:
The patent transitions from two-dimensional similarity matrix computations to multi-dimensional feature spaces through neural network transformations. By projecting point cloud features into higher-dimensional latent spaces through multiple decoding branches, the system achieves more efficient representations that reduce computational complexity while improving segmentation capability.
3Measurement precision
If traditional segmentation methods are used, then processing can be completed, but accuracy is insufficient especially for large point clouds
Solution Approach 1:
The patent performs preliminary feature extraction and encoding before segmentation by learning hierarchical representations through neural network layers. This preliminary processing organizes the large point cloud data into structured feature representations that enable accurate segmentation even when the point cloud contains a large number of points, as the complex data is pre-processed into manageable feature structures.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Disclosed are a point cloud segmentation method, a computer-readable storage medium and a computer device. The method comprises: encoding a point cloud to be processed to obtain a shared feature (S202), the shared feature referring to a feature shared at a semantic level and an instance level; decoding the shared feature according to different decoding modes to respectively obtain a semantic feature and an instance feature (S204); adapting the semantic feature to an instance feature space and then fusing same with the instance feature to obtain a semantic fused instance feature of the point cloud, the semantic fused instance feature representing an instance feature fused with a semantic feature (S206); dividing the semantic fused instance feature of the point cloud to obtain a semantic fused instance feature of each point in the point cloud (S208); and determining, according to the semantic fused instance feature of each point, an instance category to which each point belongs (210). By means of the method, the efficiency of point cloud segmentation is improved.