Adaptive Threshold Object Classification via Inter-Class Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object recognition and classification systems, particularly in open set scenarios, face challenges in accurately determining whether an object belongs to a known class or not, leading to potential misclassification and inefficiencies in multimedia content processing.
Innovation Solution
A method and apparatus that access a gallery of classes, determine distances between classes, and set adaptive thresholds for classifying test objects, where the threshold for a class is based on its distance from other classes, enabling accurate classification by comparing the distance of a test object to the threshold for each class.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed thresholds are used for object classification, then the classification process is simple and fast, but the accuracy decreases leading to false classifications
Solution Approach 1:
The patent performs preliminary computation of distances between all classes during the training phase, storing these distance metrics for later use. This preliminary action enables the system to establish data-driven thresholds before actual classification occurs, improving accuracy without adding complexity during real-time operation.
Solution Approach 2:
The patent introduces dynamic thresholds that adapt based on the distribution of training data and inter-class distances. Instead of using fixed thresholds, the system computes thresholds that reflect the actual separation between classes, allowing the classification boundary to adjust according to the specific problem domain and data characteristics.
2Measurement precision
If adaptive thresholds based on class distances are used, then classification accuracy improves, but computational complexity and processing time increase
Solution Approach 1:
The patent computes all pairwise distances between classes during the training phase and stores these pre-computed values. When a new test object needs classification, the system only needs to compare the object against the pre-stored class distances and thresholds, avoiding redundant computations and significantly reducing classification time.
Solution Approach 2:
The threshold computation is performed dynamically during training based on the actual data distribution, but once established, these thresholds remain fixed for efficient classification. This dynamic training phase followed by static deployment phase resolves the time complexity issue.
3Productivity
If simple distance comparison is used for classification, then the method is computationally efficient, but it cannot handle open set scenarios where objects may belong to unknown classes
Solution Approach 1:
The patent pre-computes the minimum distance from each class to any other class and uses this information to establish thresholds. During classification, if the distance from a test object to its nearest class exceeds the threshold, the system can correctly identify it as belonging to an unknown class, enabling open set classification while maintaining efficiency.
Solution Approach 2:
The patent introduces thresholds as an intermediary mechanism between simple distance comparison and complex classification logic. These thresholds act as a mediator that enables the system to distinguish between known and unknown classes without requiring complex models, maintaining computational efficiency while gaining open set capability.
Data Source
AI summary
In accordance with various example embodiments, methods, apparatuses, and computer program products are provided. A method comprises accessing a gallery comprising a plurality of classes, determining distances between classes of the plurality of classes, and determining thresholds for one or more classes of the gallery for classifying test objects in the classes, wherein threshold for a class is determined based on at least one distance of the class from at least one remaining class of the plurality of classes. The apparatus comprises at least one processor and at least one memory, configured to, cause the apparatus to perform accessing a gallery comprising a plurality of classes, determining distances between classes of the plurality of classes, and determining thresholds for classes of the gallery for classifying test objects in the classes, wherein threshold for a class is determined based on distances of the class from remaining classes of the gallery.