Pooling Selection Table for Scale-Aware Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural networks (CNNs) face difficulties in precisely determining bounding boxes around objects in test images due to the use of fixed pooling operations regardless of object class or scale, leading to inaccurate object detection.
Innovation Solution
A method and device for constructing a table that determines optimal pooling types based on object categories and scales, using max and average pooling operations to improve object detection accuracy by selecting the most suitable pooling type for each category and scale.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed pooling operation (max pooling or average pooling) is applied regardless of object class or scale, then the device complexity is reduced and the operation is simplified, but the object detection precision deteriorates because the pooling operation cannot adapt to different object characteristics
Solution Approach 1:
The patent implements dynamic pooling by selecting different pooling operations (max pooling, average pooling, or linear interpolation) based on the scale of the object detected. The pooling type is dynamically adjusted according to the bounding box size, allowing the system to adapt to different object scales rather than using a fixed pooling operation for all cases.
Solution Approach 2:
The patent applies different pooling operations to different regions or scales of objects within the image. Specifically, it uses max pooling for larger objects, average pooling for medium-sized objects, and linear interpolation for small objects, thereby optimizing the pooling operation for each local context rather than applying a uniform operation globally.
2Reliability
If different pooling operations are applied based on object scale and category, then the object detection precision is improved by reducing feature losses, but the device complexity increases due to the need to determine optimal pooling types for each category and scale
Solution Approach 1:
The patent pre-determines and stores the optimal pooling operation for each object category and scale combination in a lookup table or configuration. During detection, the system simply queries this pre-computed information based on the detected object's category and scale, rather than performing complex real-time analysis to determine the optimal pooling type. This preliminary preparation reduces the computational burden during actual detection.
Solution Approach 2:
The patent changes the pooling operation parameters (type of pooling) based on the scale parameter of the detected object. By adjusting the pooling type parameter according to the object scale (small, medium, large), the system optimizes feature extraction for each scale category, thereby improving detection reliability without requiring complex decision-making logic.
3Loss of information
If a single pooling operation is used for all objects, then the ease of operation is maintained and the process is simple, but the loss of information increases because the pooling operation cannot preserve features appropriate for different object scales
Solution Approach 1:
The patent implements dynamic pooling by selecting different pooling operations (max pooling, average pooling, or linear interpolation) based on the scale of the object detected. The pooling type is dynamically adjusted according to the bounding box size, allowing the system to adapt to different object scales rather than using a fixed pooling operation for all cases.
Solution Approach 2:
The patent changes the pooling operation parameters (type of pooling) based on the scale parameter of the detected object. By adjusting the pooling type parameter according to the object scale (small, medium, large), the system optimizes feature extraction for each scale category, thereby improving detection reliability without requiring complex decision-making logic.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for constructing a table including information on a pooling type based on ranges of scale of at least one object in at least one image for constructing table is provided. The method includes steps of: generating a first pooled feature map by applying max pooling and a second pooled feature map by applying avg pooling; and acquiring a first candidate bounding box by using the first pooled feature map and acquiring a second candidate bounding box by using the second pooled feature map; and comparing a first degree of similarity between the first candidate bounding box and a GT bounding box with a second degree of similarity between the second candidate bounding box and the GT bounding box to thereby construct the table so as to include information on respective optimal pooling types by respective ranges of the scale of the object.