Subpart-Guided Object Detection Filtering for Crowded Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using anchor-based neural networks face challenges in setting the overlap threshold for non-maximum suppression, leading to increased false positives or reduced recall, especially when dealing with multiple detections of the same object in crowded scenes.
Innovation Solution
A method that utilizes object subparts, such as heads or limbs, to guide the filtering process, employing two stages of filtering: a less aggressive first filtering followed by a more aggressive second filtering, based on classification confidence scores and proximity measures, to distinguish between multiple detections of the same object.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the overlap threshold for non-maximum suppression is set high, then the number of false positives increases, but the recall decreases
Solution Approach 1:
The filtering process is divided into two distinct stages: a first filtering stage that retains more proposals (less aggressive) and a second filtering stage that is more aggressive. This segmentation allows the system to first preserve potential detections and then selectively remove false positives, resolving the contradiction between precision and recall by applying different filtering strictness at different stages.
Solution Approach 2:
The patent dynamically adjusts the filtering aggressiveness between two stages. The first filtering uses a higher overlap threshold to preserve recall, while the second filtering uses a lower threshold to improve precision. This dynamic adjustment of filtering parameters allows the system to optimize both precision and recall rather than being constrained by a single fixed threshold.
2Reliability
If the overlap threshold is set low, then the recall improves, but the number of false positives increases
Solution Approach 1:
The filtering process is divided into two distinct stages: a first filtering stage that retains more proposals (less aggressive) and a second filtering stage that is more aggressive. This segmentation allows the system to first preserve potential detections and then selectively remove false positives, resolving the contradiction between precision and recall by applying different filtering strictness at different stages.
Solution Approach 2:
The first filtering stage performs a preliminary retention of object proposals with a higher overlap threshold, preserving potential true detections before the second filtering stage applies more aggressive removal. This preliminary action ensures that recall is maintained while the subsequent stage cleans up false positives, achieving both high recall and precision.
3Ease of operation
If a single filtering threshold is used, then the process is simple, but it cannot adapt to different scene types
Solution Approach 1:
The patent dynamically adjusts the filtering aggressiveness between two stages. The first filtering uses a higher overlap threshold to preserve recall, while the second filtering uses a lower threshold to improve precision. This dynamic adjustment of filtering parameters allows the system to optimize both precision and recall rather than being constrained by a single fixed threshold.
Solution Approach 2:
The filtering process is divided into two distinct stages: a first filtering stage that retains more proposals (less aggressive) and a second filtering stage that is more aggressive. This segmentation allows the system to first preserve potential detections and then selectively remove false positives, resolving the contradiction between precision and recall by applying different filtering strictness at different stages.
Data Source
AI summary
A method for object detection (post-processing) in an image is provided, and includes obtaining, from one or more artificial neural network (ANN) entities trained to localize objects and one or more subparts of such objects in images, a plurality of object proposals and one or more subpart proposals in a same image; performing a first filtering of the object proposals; matching subpart proposals with corresponding object proposals remaining after the first filtering, and performing a second filtering of the unmatched object proposals remaining after the first filtering. The first and second filtering are based on classification confidence scores and proximity scores of the object proposals, and the second filtering is statistically more aggressive than the first filtering. A corresponding device, computer program and computer program product are also provided.