3D Feature Map Decomposition for 2D Convolution Hardware
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hardware components primarily support two-dimensional convolution processing and are unable to handle three-dimensional convolution processing, which is essential for applications like automatic driving where 3D models are increasingly demanded.
Innovation Solution
A method and apparatus that decompose a three-dimensional feature map into multiple two-dimensional feature maps based on depth information and an overlay parameter, allowing two-dimensional convolution processing to simulate three-dimensional convolution, thereby enabling three-dimensional convolution processing without modifying existing hardware configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If hardware components are designed to support only two-dimensional convolution processing, then hardware complexity is reduced and manufacturing is easier, but the ability to process three-dimensional feature maps is lost
Solution Approach 1:
The three-dimensional feature map is segmented into multiple two-dimensional feature maps by dividing the depth dimension into discrete layers. Each two-dimensional feature map can be independently processed by existing hardware, while the collection of layers represents the full three-dimensional data structure. This segmentation allows standard hardware to handle 3D data without requiring 3D convolution capabilities.
Solution Approach 2:
The patent transforms the three-dimensional convolution problem into a series of two-dimensional convolution problems by introducing a layer dimension. Instead of performing convolution operations in three spatial dimensions, the method processes each two-dimensional layer separately and then aggregates the results. This dimensionality reduction enables existing 2D hardware to effectively process 3D feature maps.
2Adaptability or versatility
If hardware is upgraded to support three-dimensional convolution processing, then three-dimensional feature map processing capability is improved, but hardware complexity and configuration changes are required
Solution Approach 1:
Instead of creating new hardware specifically for three-dimensional convolution, the patent creates multiple copies of existing two-dimensional processing units. Each processing unit handles a specific two-dimensional layer of the three-dimensional feature map. By replicating proven 2D hardware designs rather than inventing new 3D hardware, the system achieves 3D processing capability while maintaining hardware simplicity and leveraging existing成熟 designs.
3Device complexity
If three-dimensional convolution processing is implemented using existing two-dimensional hardware, then hardware configuration changes are avoided, but additional processing steps are required
Solution Approach 1:
The patent maintains continuous processing by pipelining the decomposition and reconstruction operations. While one set of processing units is computing convolutions on one layer, other units simultaneously process different layers. The aggregation of results from multiple layers continues without interruption, ensuring that the additional processing steps required for 3D convolution do not create significant bottlenecks.
Data Source
AI summary
Disclosed are a feature extraction method and apparatus for a three-dimensional feature map, a storage medium, and an electronic device. The method includes: determining an overlay parameter based on depth information of a three-dimensional feature map to be processed; decomposing the three-dimensional feature map into a plurality of target two-dimensional feature maps based on the depth information and the overlay parameter; performing two-dimensional convolution processing on each of the plurality of target two-dimensional feature maps to obtain a plurality of initial feature maps; and determining a target feature map corresponding to the three-dimensional feature map based on the plurality of initial feature maps.