Octree-Based Convolutional Neural Networks for Sparse 3D Shape Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current solutions for representing three-dimensional shapes and performing convolutional neural network processing require significant storage space and computing resources due to the non-sparse representation of triangle meshes or irregularly sampled point clouds.
Innovation Solution
A convolutional neural network architecture based on an octree data structure that includes empty nodes as leaf nodes, reducing the number of nodes to store and operate on, and utilizing a shuffled key method for efficient access and operations, such as convolution, pooling, and unpooling, optimized for graphic processing units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional non-sparse representation (triangle meshes or irregularly sampled point clouds) is used for three-dimensional shapes, then complete shape information is preserved, but storage space and computing resources are significantly increased
Solution Approach 1:
The three-dimensional space is segmented into an octree data structure that divides the space into hierarchical levels of cubic regions. This segmentation allows the system to represent only the necessary portions of space containing the three-dimensional shape, rather than storing the entire space or using complex mesh structures. The octree recursively subdivides space into eight octants per node, enabling efficient spatial partitioning that reduces storage requirements while preserving shape information.
Solution Approach 2:
The patent applies different representation qualities to different regions of the three-dimensional space. Regions containing the three-dimensional shape are represented with higher detail (non-empty nodes), while empty regions are represented with lower detail (empty nodes). This local differentiation allows the system to allocate computational resources and storage space efficiently, focusing detail where needed and using compact representations where unnecessary, thereby reducing overall storage space without losing critical shape information.
2Productivity
If traditional non-sparse representation is used for three-dimensional shapes, then accurate shape processing is achieved, but computational resources and processing time are significantly increased
Solution Approach 1:
The octree structure segments the three-dimensional space into a hierarchical organization of nodes, allowing convolutional neural network operations to be performed only on relevant non-empty nodes rather than processing the entire space uniformly. This segmentation enables the system to skip over empty regions entirely, significantly reducing the number of computations required while maintaining accurate shape processing on the regions that actually contain data.
Solution Approach 2:
The patent applies convolutional operations selectively to only the necessary portions of the three-dimensional shape data (non-empty nodes in the octree), rather than performing operations on the complete three-dimensional space. This partial action approach computes only where needed, reducing computational resource consumption and processing time while still achieving accurate shape analysis and recognition tasks.
3Quantity of substance
If sparse octree representation is used to reduce storage space, then storage efficiency is improved, but access and operation efficiency may be degraded
Solution Approach 1:
The patent pre-organizes the octree data structure and pre-computes spatial relationships between nodes before performing convolutional operations. The shuffled key method pre-arranges node indexing to optimize memory access patterns, and the hierarchical structure pre-establishes parent-child relationships that facilitate efficient traversal. This preliminary organization enables the system to maintain fast access efficiency despite the sparse representation, as the data is structured for optimal retrieval rather than requiring complex on-the-fly computations.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The implementations of the subject matter described herein relate to an octree-based convolutional neural network. In some implementations, there is provided a computer-implemented method for processing a three-dimensional shape. The method comprises obtaining an octree for representing the three-dimensional shape. Nodes of the octree include empty nodes and non-empty nodes. The empty nodes exclude the three-dimensional shape and are leaf nodes of the octree, and the non-empty nodes include at least a part of the three-dimensional shape. The method further comprises for nodes in the octree with a depth associated with a convolutional layer of a convolutional neural network, performing a convolutional operation of the convolutional layer to obtain an output of the convolutional layer.