Ambiguous Object Classification via Multi-Class and Single-Stage Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object classification systems are generalist and prone to misclassifying objects due to multiple classes, leading to reduced accuracy and confidence in edge analytics applications, especially with limited compute power.
Innovation Solution
A method involving a multi-class object classification model that outputs confidence scores, identifies candidate classes within a threshold range, and employs single-class object classification models for each candidate class to determine the true class of ambiguous objects, improving classification accuracy while managing computation resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a multi-class object classification model is used to detect multiple object classes, then the system can identify various objects (cars, buses, people, bicycles), but the classification accuracy deteriorates due to class confusion (e.g., misclassifying a car as a train)
Solution Approach 1:
The patent segments the classification process into two distinct stages: first, a multi-class OCM performs initial classification across multiple object categories; second, single-class OCMs are selectively applied to candidate classes that have confidence scores within a threshold range. This segmentation allows the system to maintain broad detection capability while achieving high precision for ambiguous objects by applying specialized classifiers only where needed.
2Measurement precision
If multiple single-class object classification models are executed for each candidate class to improve accuracy, then classification precision improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies partial action by executing single-class OCMs only for candidate classes whose confidence scores fall within a threshold range of the highest confidence score, rather than running all possible single-class models. This selective approach achieves high classification accuracy for ambiguous objects while avoiding the computational burden of evaluating every possible class combination, thus balancing precision with resource constraints in edge analytics environments.
3Productivity
If conventional multi-class object classification models are used, then the system can process images quickly, but the reliability of classification results deteriorates due to misclassification of ambiguous objects
Solution Approach 1:
The patent implements a feedback mechanism where the multi-class OCM first provides initial classification results with confidence scores, and objects with ambiguous classifications (confidence scores within threshold range) are identified and fed back for re-evaluation by single-class OCMs. This feedback loop ensures that only uncertain cases receive additional processing, maintaining high productivity while improving reliability for ambiguous objects through iterative refinement.
Data Source
AI summary
Example implementations include a method, apparatus and computer-readable medium for object classification, including executing a multi-class object classification model (OCM) on an input image depicting an object, wherein the multi-class OCM is configured to output, for each respective class of a plurality of classes, a respective confidence score indicative of a likelihood of the object being of a member the respective class. The implementations further include detecting a plurality of confidence scores outputted by the multi-class OCM that are within a threshold range, wherein classes associated with the plurality of confidence scores are candidate classes of the object. Additionally, the implementations further include executing, on the input image, a single-class OCM for each of the candidate classes and outputting a final class for the object in the input image based on a result of each single-class OCM.