Object Detection Filtering Using Subpart-Guided Proposal Matching
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 either increased false positives or reduced recall, particularly 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 based on higher proximity thresholds and a more aggressive second filtering to distinguish between object proposals, reducing the risk of discarding separate objects close together.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the overlap threshold is set too high, then the risk of retaining several boxes indicative of the same object increases, but the precision is reduced
Solution Approach 1:
The patent segments the object detection problem into two stages: first filtering based on overlap thresholds, then second filtering based on subpart proposal matching. This segmentation allows each stage to address specific aspects of the contradiction - the first stage reduces false positives through overlap filtering, while the second stage preserves recall by identifying objects with distinctive subparts even when overlap is high.
Solution Approach 2:
The patent introduces subpart proposals as an intermediary element between object proposals and final detection results. These subpart proposals (e.g., heads, limbs for persons; license plates for vehicles) serve as mediators that provide additional discriminative information to resolve ambiguous cases where overlap-based filtering alone would either retain false positives or discard valid detections.
2Measurement precision
If the overlap threshold is set too low, then the risk of discarding boxes indicative of different objects increases, but the recall is reduced
Solution Approach 1:
The patent performs preliminary action by generating subpart proposals before the filtering stages. These subpart proposals are prepared in advance and used during both filtering stages to guide the retention or discarding of object proposals. This preliminary preparation of subpart information enables the system to make more informed decisions about preserving detections that might otherwise be discarded due to low overlap thresholds.
3Device complexity
If a single filtering stage is used, then the processing is simpler, but the ability to distinguish between closely positioned objects is reduced
Solution Approach 1:
The patent divides the filtering process into two distinct stages: first filtering based on overlap thresholds to reduce false positives, and second filtering based on subpart proposal matching to preserve detections of closely positioned objects. This segmentation of the filtering process allows each stage to specialize in different aspects of the detection quality trade-off, achieving better overall performance than a single unified filtering stage.
Data Source
Figure 1A~1B
Figure 1C~1D
Figure 2~3A
AI summary
A method (200) for object detection (post-processing) in an image is provided, and includes obtaining (S210), 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 (S220) a first filtering of the object proposals; matching (S230) subpart proposals with corresponding object proposals remaining after the first filtering, and performing (S240) 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.