Tag-Based NMS for Nearby Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolutional neural networks (CNN) face challenges in accurately detecting objects that are close to each other, as they often lower the certainty factors of object rectangles, leading to incorrect deletion of detected objects.
Innovation Solution
The method separates object rectangles with the same category into groups based on object tags, performing non-maximum suppression (NMS) processing within each group to maintain certainty factors, ensuring accurate detection of objects close to each other.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional NMS processing is performed to eliminate duplicate object rectangles, then processing efficiency is improved, but objects that are close to each other may be incorrectly deleted due to excessive overlap
Solution Approach 1:
The patent changes the parameter used for grouping from category alone to a combination of category and object tag (embedding vector). This allows objects of the same category to be separated into different groups based on their semantic similarity, preventing incorrect deletion of close objects while maintaining processing efficiency through grouped NMS operations.
2Reliability
If Soft-NMS processing is used to lower certainty factors for overlapping objects, then detection of close objects is improved, but certainty factors may decrease without limit causing accurate objects to be lost
Solution Approach 1:
The patent segments object rectangles into different groups based on category and object tag. By performing NMS processing separately for each group, the patent avoids the need to lower certainty factors across all overlapping objects, preserving the original certainty information while still eliminating duplicates within each semantic group.
3Device complexity
If NMS processing is performed on all object rectangles of the same category, then duplicate detection is reduced, but objects with high overlap are incorrectly eliminated even when they represent distinct targets
Solution Approach 1:
The patent applies different treatment to different groups of object rectangles based on their object tags. Objects with similar semantic meanings (high embedding similarity) are grouped together and subjected to NMS, while objects with different semantics are kept separate. This local differentiation preserves distinct objects that happen to have spatial overlap.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
An object detection device (1) includes an object detection unit (2) that detects an object from an image including the object by neural computation using a CNN. The object detection unit (2) includes: a feature amount extraction unit (2a) that extracts a feature amount of the object from the image; an information acquisition unit (2b) that obtains a plurality of object rectangles indicating candidates for the position of the object on the basis of the feature amount and obtains information and a certainty factor of a category of the object for each of the object rectangles; and an object tag calculation unit (2c) that calculates, for each of the object rectangles, an object tag indicating which object in the image the object rectangle is linked to, on the basis of the feature amount. The object detection device (2) further includes an excess rectangle suppression unit (4) that separates a plurality of object rectangles for which a category of the object is the same into a plurality of groups according to the object tags, and deletes an excess object rectangle in each of the separated groups on the basis of the certainty factor.