Anchor Box Parameter Determination for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems require manual setting of hyperparameters such as scales, aspect ratios, and the number of anchor boxes, leading to inefficiencies and unused anchor boxes.
Innovation Solution
A parameter determination apparatus and method that automatically determine the parameters of anchor boxes by generating distribution information, clustering it, and determining anchor box parameters based on the clusters, thereby reducing the load of object detection processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a countless number of anchor boxes are scattered throughout the image with various sizes and aspect ratios, then object detection coverage is improved, but processing load increases and many anchor boxes remain unused
Solution Approach 1:
The patent extracts only the necessary anchor boxes from the set of all possible anchor boxes. By analyzing the distribution of ground truth bounding boxes and selecting anchor boxes that actually correspond to objects in the training data, it removes unnecessary anchor boxes that would increase processing load without contributing to detection accuracy.
Solution Approach 2:
The patent changes the parameters of anchor boxes (sizes and aspect ratios) based on the statistical distribution of objects in the training data. Instead of using a fixed set of anchor box parameters, it dynamically adjusts them to match the actual object characteristics, ensuring that only relevant anchor boxes are used.
2Adaptability or versatility
If manual setting of hyperparameters such as scales, aspect ratios, and number of anchor boxes is performed, then flexibility in adjustment is improved, but time consumption and inefficiency increase
Solution Approach 1:
The patent enables the system to automatically determine anchor box parameters by itself, without requiring manual intervention. It analyzes the training data distribution and autonomously selects appropriate anchor box sizes, aspect ratios, and quantities, making the system self-configuring and eliminating time-consuming manual hyperparameter tuning.
Solution Approach 2:
The patent performs preliminary analysis of the training data distribution before actual object detection. By pre-computing the optimal anchor box parameters from the training data statistics, it prepares the system in advance, so that during detection phase, no additional manual adjustment is needed.
3Measurement precision
If anchor boxes are set to correspond to objects having various shapes, then detection accuracy is improved, but many anchor boxes are not actually used for learning
Solution Approach 1:
The patent extracts only those anchor boxes that correspond to actual objects in the training data. By filtering out anchor boxes that do not match any ground truth bounding box, it ensures that every anchor box used in learning is meaningful and contributes to detection accuracy, eliminating wasted computational resources on unused anchor boxes.
Solution Approach 2:
The patent adjusts anchor box parameters (sizes and aspect ratios) to match the actual distribution of objects in the training data. This ensures that anchor boxes are tailored to the specific characteristics of the target objects, improving detection accuracy while maintaining learning efficiency by avoiding mismatched anchor boxes.
Data Source
AI summary
A detection object analysis unit (4) is a parameter determination apparatus that determines parameters of a plurality of anchor boxes to be used in a sliding window method when a bounding box and a class of an object in an image are detected using a neural network and the sliding window method. The detection object analysis unit (4) includes a distribution generation unit (11) that generates distribution information of parameters of bounding boxes indicated by object specifying information of a plurality of pieces of learning data. The detection object analysis unit (4) includes a clustering processing unit (12) that generates a plurality of clusters by clustering the distribution information. The detection object analysis unit (4) includes a parameter determination unit (13) that determines the parameters of the plurality of anchor boxes based on the plurality of clusters.