Parallel Non-Maximum Suppression for Redundant Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing non-maximum suppression (NMS) algorithms for object detection tasks in neural networks cannot fully parallelize the suppression of redundant bounding boxes due to their reliance on sorting and handling bounding boxes in a descending order, limiting their performance.
Innovation Solution
The proposed method parallelizes NMS by performing multiple suppression processes in parallel, constraining each process to an area surrounding a candidate bounding box and comparing only neighboring boxes within this area, thereby eliminating the need for sorting and enabling full parallelization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional NMS algorithms sort and process bounding boxes in descending order, then suppression accuracy is maintained, but processing speed and parallelization capability deteriorate
Solution Approach 1:
The patent divides the bounding box suppression task into multiple independent regions or channels, allowing parallel processing of different segments simultaneously. Each segment processes a subset of bounding boxes independently, eliminating the need for global sorting while maintaining local suppression accuracy.
Solution Approach 2:
The patent introduces a spatial or channel dimension to the suppression process by organizing bounding boxes into multiple parallel processing paths. This dimensional transformation allows the system to process bounding boxes in parallel across different dimensions rather than sequentially in a single sorted list.
2Reliability
If traditional NMS algorithms process all bounding boxes sequentially, then comprehensive suppression is achieved, but computational overhead and processing time increase
Solution Approach 1:
The patent segments the bounding box set into multiple groups that can be processed in parallel. By dividing the comprehensive suppression task into independent segments, the system achieves both completeness (all boxes are suppressed) and efficiency (parallel processing reduces total time).
Solution Approach 2:
The patent merges multiple parallel suppression processes into a unified output. Different parallel processing paths work simultaneously on different subsets of bounding boxes, then their results are combined to achieve comprehensive suppression across all bounding boxes.
3Measurement precision
If NMS algorithms use global sorting of bounding boxes, then optimal suppression decisions are made, but device complexity and computational resources increase
Solution Approach 1:
The patent applies local quality by making suppression decisions based on local relationships within segments rather than global sorting. Each segment makes optimal suppression decisions based on its local bounding box characteristics, reducing the overall algorithmic complexity while maintaining decision quality.
Data Source
AI summary
Apparatuses, systems, and techniques to perform non-maximum suppression (NMS) in parallel to remove redundant bounding boxes. In at least one embodiment, two or more parallel circuits to perform two or more portions of a NMS algorithm in parallel to remove one or more redundant bounding boxes corresponding to one or more objects within one or more digital images.