Multi-Scale Likelihood Maps for Overlapping Object Position Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object position estimation methods struggle with processing speed limitations and accuracy when objects overlap in an image, making it difficult to accurately estimate the position of each object.
Innovation Solution
An object position estimation device that utilizes a first and second feature extraction unit to generate feature maps through convolution processing, followed by likelihood map estimation units to determine the probability of object presence at different sizes, allowing for robust and accurate estimation of object positions even when objects overlap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the estimator scans the image by continuously changing the position and size of partial regions, then the object position estimation accuracy is improved, but the processing speed deteriorates due to computer processing limitations
Solution Approach 1:
The patent segments the object detection task by dividing the feature extraction process into multiple stages (first feature map, second feature map with different resolutions). This allows parallel processing of different feature representations, improving processing speed while maintaining the ability to accurately detect objects at various positions and sizes through multiple likelihood maps.
2Productivity
If the estimator uses a single feature map for object detection, then the processing speed is maintained, but the accuracy deteriorates when objects overlap or are shielded by other objects
Solution Approach 1:
The patent adds a dimensional aspect by creating feature maps at different resolutions (first feature map with higher resolution, second feature map with lower resolution). This multi-resolution approach enables the system to detect objects effectively whether they are small, large, partially visible, or overlapping, thereby improving accuracy without sacrificing processing speed through efficient parallel processing.
3Measurement precision
If the system performs comprehensive convolution processing on the target image, then the feature extraction accuracy is improved, but the processing time increases
Solution Approach 1:
The patent performs preliminary convolution processing to generate the first feature map, then uses this intermediate result to generate the second feature map through additional convolution. This staged approach allows the system to maintain high feature extraction accuracy while reducing overall processing time by reusing intermediate results and enabling parallel processing of likelihood map generation from multiple feature maps.
Data Source
AI summary
An object position estimation device (1) is provided with: a feature extraction unit (10) including a first feature extraction unit (21) which generates a first feature map by subjecting a target image to a convolution computation process, and a second feature extraction unit (22) which generates a second feature map by also subjecting the first feature map to the convolution computation process; and a likelihood map estimation unit (20) including a first position likelihood estimation unit (23) which, by using the first feature map, estimates a first likelihood map indicating the probability that first objects having a first size are present in the target image, and a second position likelihood estimation unit (24) which, by using the second feature map, estimates a second likelihood map indicating the probability that second objects having a second size, which is greater than the first size, are present in the target image.