Bit-Reduced Radix Sorting for Accelerated Non-Maximum Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network object detection tasks face inefficiencies in processing redundant bounding boxes due to time-consuming sorting operations that utilize the full bits of floating-point confidence scores, particularly in algorithms like NMS.
Innovation Solution
Implementing a bit-reduced radix sorting method that transforms confidence scores into a specified range with a shared exponent field, allowing for faster sorting and suppression of redundant bounding boxes by using only the mantissa field for comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-bit floating-point confidence scores are used for sorting in NMS algorithms, then measurement precision is maintained, but processing time increases significantly
Solution Approach 1:
The patent extracts only the necessary mantissa bits from the full floating-point confidence scores for sorting operations. By separating the mantissa field from the exponent field and using only the mantissa for comparisons, the method reduces the data width processed during sorting while preserving the precision needed to distinguish between different confidence scores.
Solution Approach 2:
The patent applies partial action by using only a portion (the mantissa field) of the complete confidence score data for sorting purposes. This partial use of data is sufficient to achieve the required sorting precision while significantly reducing the computational burden compared to processing the full floating-point representation.
2Reliability
If traditional sorting algorithms are used for bounding boxes, then accurate suppression is achieved, but computational complexity increases
Solution Approach 1:
The patent changes the parameter used for sorting from full floating-point confidence scores to reduced-precision mantissa values. This parameter transformation maintains the relative ordering of confidence scores while reducing the computational complexity of sorting operations, as smaller data widths require fewer processing cycles and less memory bandwidth.
Solution Approach 2:
The patent creates a simplified copy of the confidence score data by extracting only the mantissa field. This copied representation is then used for sorting operations instead of the original full-precision values, achieving the same sorting result with reduced computational complexity.
Data Source
AI summary
Apparatuses, systems, and techniques to perform non-maximum suppression (NMS) with a bit-reduced radix sort to remove redundant bounding boxes are described. In at least one embodiment, one or more circuits perform i) a bit-reduced radix sort operation to sort a list of confidence scores associated with a set of bounding boxes corresponding to one or more objects within one or more digital images and ii) a non-maximum suppression (NMS) operation on the sorted list to remove one or more redundant bounding boxes from the set.