Object Count Estimation Using Dual Feature Maps for Size-Specific Ranges
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies for estimating the number of objects in an image require a significant amount of labor to set up partial regions, particularly when determining head sizes and camera parameters.
Innovation Solution
An object count estimation apparatus that uses a feature extraction network to generate two feature maps, which are then processed by counting networks to estimate the number of objects within specific size ranges for different image regions, reducing the need for manual setting of partial regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual setting of partial regions is performed to ensure accurate object counting, then counting precision is improved, but labor requirements increase significantly
Solution Approach 1:
The image is divided into multiple partial regions automatically based on detected object positions and sizes. The segmentation is performed dynamically without manual intervention, where each detected object defines its own partial region for counting, thus eliminating the need for manual partial region setting while maintaining counting accuracy.
Solution Approach 2:
The system performs automatic partial region setting by utilizing the detected object information itself. The object detector identifies objects and their positions, and this information is directly used to define the partial regions for counting, making the system self-sufficient without requiring external manual configuration.
2Extent of automation
If camera parameters are estimated accurately to compute head sizes automatically, then partial region setting automation is achieved, but labor for parameter estimation increases
Solution Approach 1:
The system extracts only the necessary information (object positions and sizes) directly from the image using an object detector, rather than estimating camera parameters. This extraction approach eliminates the time-consuming camera parameter estimation process while still enabling automatic partial region computation.
Solution Approach 2:
The approach changes from using camera parameters to using directly observable object parameters (position and size) from the image. This parameter substitution eliminates the need for camera calibration and estimation, achieving automation without the associated time cost.
3Device complexity
If a single counting network is used for all object sizes, then device complexity is reduced, but counting precision for varying sizes deteriorates
Solution Approach 1:
The counting task is segmented into multiple size ranges, with each counting network specialized for a specific size range. This segmentation allows each network to focus on particular object sizes, improving precision without requiring an overly complex universal network.
Solution Approach 2:
Instead of one network attempting to handle all sizes, multiple networks with specialized size ranges are deployed. This partial specialization approach improves overall counting accuracy by having each network excel at its specific size range rather than being mediocre at all sizes.
Data Source
AI summary
An object count estimation apparatus (2000) includes a first feature extraction network (2042), a first counting network (2044), a second feature extraction network (2062), and a second counting network (2064). The first feature extraction network (2042) generates a first feature map (20) by performing convolution processing on a target image (10). The first counting network (2044) estimates the number of target objects having a size included in a first predetermined range by performing processing on the first feature map (20). The second feature extraction network (2062) generates a second feature map (30) by performing convolution processing on the first feature map (20). The second existence estimation network (2064) estimates the number of target objects having a size included in a second predetermined range by performing processing on the second feature map (30). A size included in the first predetermined range is smaller than a size included in the second predetermined range.