Proposal Processing Method for Object Detection Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing object detection systems, such as faster R-CNN, face challenges in achieving real-time performance due to the time-consuming non-maximum suppression (NMS) algorithm, which is inefficient in processing large quantities of overlapping region proposals, particularly in contexts requiring high real-time performance like autonomous driving.
Innovation Solution
A proposal processing method that determines a score threshold based on statistical properties of proposal scores, selects proposals with scores higher than the threshold, and uses a hardware register array for unsorted suppression to efficiently obtain non-overlapping proposals with high scores, reducing time complexity and improving processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If non-maximum suppression (NMS) algorithm is used to process region proposals, then non-overlapping proposals with high scores can be obtained, but processing time increases significantly
Solution Approach 1:
The patent segments the NMS process into two distinct stages: (1) score threshold filtering to select candidate proposals, and (2) IoU-based suppression to remove overlapping proposals. This segmentation allows each stage to operate independently and efficiently, reducing overall processing time while maintaining filtering accuracy.
Solution Approach 2:
The patent applies preliminary score threshold filtering before performing the more computationally intensive IoU-based suppression. By pre-filtering proposals based on their scores (using statistics such as mean and standard deviation to determine thresholds), the system reduces the number of proposals that need to undergo full suppression processing, thereby significantly reducing processing time.
2Reliability
If traditional NMS algorithm is used to find non-overlapping proposals, then detection accuracy is maintained, but real-time performance is compromised
Solution Approach 1:
The patent changes the parameter used for proposal selection from traditional sequential IoU-based filtering to a statistics-based score threshold approach. By using statistical parameters (mean, standard deviation) to dynamically determine score thresholds, the system maintains detection accuracy while enabling real-time processing speeds suitable for applications like autonomous driving.
3Quantity of substance
If all proposals are processed through complete NMS, then no proposals are lost, but computational complexity increases
Solution Approach 1:
The patent extracts and removes proposals that fail to meet the score threshold criterion before they enter the main suppression process. This extraction step eliminates a large portion of low-scoring proposals that would otherwise consume computational resources in the suppression algorithm, thereby reducing overall computational complexity while preserving the quality of final output proposals.
Data Source
AI summary
Provided are a proposals processing method and related products for object detection, where the method comprising: determining a score threshold according to a statistics property of scores of a first quantity of proposal; selecting a second quantity of proposals with a score higher than the score threshold from the first quantity of proposals; and suppressing to the second quantity of proposals to get a third quantity of non-overlapping proposals with high scores, as such, time complexity for selecting a second quantity of proposals is reduced, and efficiency for selection of the proposals to be used for suppression is improved, thereby improving proposal processing efficiency and object detection efficiency.