Object Detection Box Selection for Closely Spaced Targets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection methods using non-maximum suppression suffer from error suppression when objects of the same category are closely located, as the intersection over union (IOU) between detection boxes is high, leading to inaccurate identification.
Innovation Solution
Determine a target prediction region by considering both the overlapping area and similarity between objects in prediction regions, using a coverage region and similarity measure to refine the selection process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If non-maximum suppression is used to select target detection boxes, then processing efficiency is improved, but identification accuracy deteriorates when objects are closely located
Solution Approach 1:
The patent changes the suppression criterion from IOU (intersection over union) to area of coverage region. By using the area of the coverage region (jointly covered area between prediction regions) instead of the ratio-based IOU metric, the method adapts the suppression parameter to better handle closely located objects, preventing erroneous suppression while maintaining processing efficiency
Solution Approach 2:
The patent introduces a coverage region as an intermediary concept between the prediction regions. This coverage region represents the jointly covered area and serves as a mediator to determine suppression decisions. By calculating the area of this intermediate coverage region and comparing it with thresholds, the method achieves more accurate suppression decisions for closely located objects
2Ease of operation
If IOU threshold is used for suppression, then simple decision making is achieved, but error suppression occurs for closely located objects
Solution Approach 1:
The patent changes the suppression parameter from IOU (a ratio metric) to the absolute area of the coverage region. This parameter change allows for more reliable suppression decisions because the area metric better reflects the actual spatial relationship between closely located objects, reducing error suppression while maintaining operational simplicity through threshold comparison
3Productivity
If detection boxes with high overlapping area are suppressed, then redundant detections are reduced, but valid detection boxes are erroneously removed
Solution Approach 1:
The patent replaces the mechanical IOU ratio calculation with an area-based coverage region metric. By substituting the ratio-based suppression mechanism with an area-based one, the method better preserves valid detection boxes for closely located objects while still removing redundant detections, improving both detection quality and reliability
Solution Approach 2:
The patent creates a coverage region as a copy or representation of the jointly covered space between prediction regions. This coverage region copy is then used for suppression decisions instead of directly using the original prediction regions, allowing for more accurate determination of which boxes to suppress
Data Source
AI summary
This disclose relates to a data processing method and apparatus. The method includes: acquiring a first prediction region in a target image, the first prediction region being a prediction region corresponding to a maximum prediction category probability in N prediction regions in the target image, a prediction category probability being a probability that an object in a prediction region belongs to a prediction object category; determining a coverage region jointly covered by a second prediction region and the first prediction region; the second prediction region being a prediction region other than the first prediction region in the N prediction regions; and determining a target prediction region in the prediction regions based on an area of the coverage region and a similarity associated with the second prediction region, the similarity being for indicating a similarity between an object in the second prediction region and an object in the first prediction region.