Dynamic Anchor Box Selection for Neural Network Training Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing object detection methods using anchor boxes face instability in training due to position dependence and the limitation of assigning a constant maximum number of anchor boxes, leading to failed training and decreased detection reliability.
Innovation Solution
A training device and method that dynamically adjusts the number of anchor boxes assigned to a correct answer region based on similarity thresholds, allowing for flexible selection and reassignment of anchor boxes to improve training accuracy and reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a constant maximum number of anchor boxes is assigned to each correct answer label, then the assignment process is simple and fast, but training stability deteriorates when the number of suitable anchor boxes exceeds the limit
Solution Approach 1:
The patent applies dynamics by making the maximum number of anchor boxes assignable to a correct answer label variable rather than constant. The selection unit dynamically adjusts this number based on the degree of similarity between anchor boxes and correct answer labels, allowing the system to adapt to different detection scenarios and maintain training stability while preserving assignment efficiency.
2Ease of operation
If anchor boxes are assigned based on fixed similarity threshold only, then the assignment criterion is clear and simple, but position dependence causes training instability for anchor boxes far from correct answer labels
Solution Approach 1:
The patent applies parameter changes by introducing multiple criteria beyond fixed similarity threshold. The selection unit considers both the degree of similarity and the positional relationship between anchor boxes and correct answer labels. This multi-parameter approach maintains clear assignment criteria while eliminating position dependence, ensuring stable training across different spatial configurations.
3Reliability
If multiple anchor boxes are assigned to one correct answer label to solve position dependence, then training stability improves, but the fixed maximum number limit causes assignment failures when too many anchor boxes are suitable
Solution Approach 1:
The patent applies dynamics by making the maximum number of assignable anchor boxes flexible rather than fixed. The selection unit determines this number dynamically based on the specific detection context, including similarity degrees and positional relationships. This allows the system to assign an appropriate number of anchor boxes for each correct answer label, maintaining both training stability and assignment flexibility without arbitrary limits.
Data Source
AI summary
A training device acquires a degree of similarity between a correct answer region indicating a region of an object in an image and each of a plurality of anchor boxes set in advance in an image, selects, among the plurality of anchor boxes, an anchor box for which the degree of similarity is greater than or equal to a predetermined threshold, for the correct answer region, and performs training of a neural network model for detecting the object, based on the correct answer region and the selected anchor box. If a maximum number of anchor boxes have been selected for the correct answer region, the training device changes the maximum number of the anchor boxes for the correct answer region based on the degrees of similarity acquired for the anchor boxes selected for the correct answer region.