Hierarchical Softmax for Noisy Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems are inadequate for large, complex datasets with hierarchical label relationships, class distribution imbalances, and missing annotations, as they assume mutually exclusive class labels and lack mechanisms to handle these complexities effectively.
Innovation Solution
The implementation of a hierarchical softmax function within a convolutional neural network (CNN) that applies multiple softmax functions based on a class hierarchy, combined with class balancing and knowledge distillation methods to address label imbalances and recover missing annotations, enabling more accurate object detection in complex datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional object detection algorithms are used with large, complex datasets, then the dataset size and category number increase, but the algorithms cannot handle class distribution imbalance and hierarchical label relationships effectively
Solution Approach 1:
The patent segments the classification task into multiple hierarchical levels using a tree structure. Instead of treating all categories as flat and independent, the algorithm divides them into parent categories and child categories, processing them in a hierarchical manner. This segmentation allows the system to handle large datasets with many categories by breaking down the complex classification problem into smaller, manageable sub-problems at each hierarchy level.
Solution Approach 2:
The patent changes the parameter representation by introducing hierarchical labels with parent-child relationships. Rather than using traditional flat class labels, the system represents categories as nodes in a hierarchy tree, where each node has associated parameters including its parent node and child nodes. This parameter transformation enables the algorithm to capture semantic relationships and handle class distribution imbalance more effectively.
2Adaptability or versatility
If the number of label categories increases to cover more object types, then category coverage improves, but complex hierarchical label relationships create challenges for detectors assuming disjoint non-overlapping categories
Solution Approach 1:
The patent segments the label space into hierarchical levels where parent categories represent broader classes and child categories represent more specific subclasses. This segmentation transforms the complex hierarchical relationship into a structured tree format that the algorithm can process systematically, moving from general to specific categories through multiple processing stages.
Solution Approach 2:
The patent adds a hierarchical dimension to the traditional flat classification space. By organizing categories into a tree structure with multiple levels, the system transforms a one-dimensional flat label space into a multi-dimensional hierarchical space, enabling the detector to capture both broad category information and specific subclass details simultaneously.
3Loss of information
If exhaustive annotation of all object categories is attempted, then label completeness improves, but the effort and cost increase significantly resulting in missing box annotations
Solution Approach 1:
The patent applies partial action by not requiring exhaustive annotation of all possible categories at all levels. Instead, the hierarchical structure allows the system to work with partially annotated data where only certain parent categories and their relevant child categories need to be labeled. The hierarchy provides implicit information that compensates for missing annotations, reducing the overall annotation burden while maintaining detection capability.
Solution Approach 2:
The hierarchical label structure acts as an intermediary that connects sparsely annotated data with comprehensive category coverage. Parent categories serve as intermediaries that can represent multiple child categories, allowing the system to infer the presence of specific objects even when detailed annotations are missing. This intermediary structure bridges the gap between limited annotation effort and complete category representation.
4Ease of operation
If traditional softmax function is used for classification, then implementation is simple, but it cannot handle hierarchical label relationships and assumes mutually exclusive classes
Solution Approach 1:
The patent segments the traditional single softmax function into multiple hierarchical softmax functions, one for each level of the category hierarchy. Each softmax layer processes a specific level of the hierarchy, transforming the classification task from a single-step operation into a multi-stage process that respects the hierarchical structure of labels while maintaining the mathematical simplicity of the softmax function at each stage.
Data Source
AI summary
Systems and methods are described for object detection within a digital image using a hierarchical softmax function. The method may include applying a first softmax function of a softmax hierarchy on a digital image based on a first set of object classes that are children of a root node of a class hierarchy, then apply a second (and subsequent) softmax functions to the digital image based on a second (and subsequent) set of object classes, where the second (and subsequent) object classes are children nodes of an object class from the first (or parent) object classes. The methods may then include generating an object recognition output using a convolutional neural network (CNN) based at least in part on applying the first and second (and subsequent) softmax functions. In some cases, the hierarchical softmax function is the loss function for the CNN.