3D Point Cloud Segmentation Using Uniform Grid Feature Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for 3D instance segmentation of point clouds are inefficient and dependent on 2D image processing, limiting their performance due to voxelization and pair-wise distance computations, which are impractical for large point clouds.
Innovation Solution
A deep learning-based system, Mask-MCNet, directly processes point clouds without voxelization, using a feature extraction network, object proposal network, and object classification network to generate accurate and efficient object detection and segmentation by transforming point cloud features into a uniform 3D grid and applying Monte Carlo convolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If 2D image processing and voxelization are used for 3D object detection, then the processing framework is simplified, but the handling of non-uniform point cloud data is limited and fine-detail geometrical information is degraded
Solution Approach 1:
The patent introduces an intermediary uniform 3D grid structure that mediates between the non-uniform point cloud data and the neural network processing. The grid acts as a bridge, allowing the network to process point cloud data directly without voxelization while maintaining a structured processing framework. This resolves the contradiction by preserving geometrical information through direct point cloud processing while keeping the framework simplified through the uniform grid abstraction.
2Ease of operation
If voxelization is applied to point cloud data, then the data is transformed into a uniform grid structure, but fine-detail geometrical information is lost
Solution Approach 1:
Instead of transforming the original point cloud data into voxels (which loses information), the patent creates a parallel uniform 3D grid structure that copies the spatial organization information without altering the original point cloud. The neural network processes both the original point cloud and the grid-based features simultaneously, allowing uniform structure benefits without information loss from the original geometrical details.
3Manufacturing precision
If deep neural networks process point clouds directly, then fine-detail geometrical information is preserved, but the processing complexity increases
Solution Approach 1:
The patent segments the processing task into multiple specialized neural network components: a feature extraction network that processes point cloud data, a Monte Carlo convolutional network that handles the uniform 3D grid features, and an object classification network. This segmentation reduces overall complexity by dividing the complex processing into manageable, specialized modules that can be processed independently and then integrated.
4Productivity
If Monte Carlo convolution is used for feature transformation, then processing speed is improved, but computational complexity per operation increases
Solution Approach 1:
The patent changes the parameter of the convolution operation from deterministic to probabilistic (Monte Carlo). Instead of fixed grid sampling, the system uses random sampling with probability distributions to transform features. This parameter change allows parallel computation of multiple samples, improving processing speed through hardware parallelization while the probabilistic nature simplifies the computational graph compared to deterministic multi-step operations.
Data Source
AI summary
A method of object detection in a point cloud includes: determining first features associated with points of a point cloud representing one or more objects in at least a 3D space and defining geometrical information for each point of the point cloud, a first type of network being configured to receive points of the point cloud as input; determining second point cloud features based on the first features, the second features defining local geometrical information about the point cloud at positions of nodes of a uniform 3D grid; generating an object, an object proposal defining a 3D bounding box, the 3D bounding box that may define an object; and determining, by a third type of deep neural network, a score for a 3D anchor indicating a probability that the 3D anchor, the determining being based on second features that are located in the 3D anchor.