Gesture Recognition Weighting for Imbalanced Class Frequency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current gesture recognition algorithms, based on supervised learning, often misclassify less frequent gestures due to imbalanced class frequency distributions, leading to sub-optimal recognition performance for rare classes.
Innovation Solution
A method that uses a weighting factor to balance the influence of classes during training, where frequently occurring classes are weighted less and less frequent classes are weighted more, to prevent over-training on dominant classes and improve recognition accuracy across all classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional supervised learning methods are used for gesture recognition, then frequently occurring gestures are recognized with high accuracy, but less frequent gestures are misclassified due to imbalanced class frequency distribution
Solution Approach 1:
The patent applies parameter changes by modifying the loss function to include class frequency-based weighting factors. The loss function is transformed from a standard form to a weighted form where each class contribution is scaled by its frequency distribution, thereby changing the optimization parameters to account for class imbalance and improve recognition accuracy for rare gestures
Solution Approach 2:
The patent implements partial action by selectively adjusting the weighting of specific classes in the loss function based on their frequency distribution. Instead of treating all classes equally, the method applies differential weighting where underrepresented classes receive higher weights, allowing the system to focus computational effort on improving recognition for rare gestures without completely sacrificing performance on frequent gestures
2Productivity
If the machine learner is optimized using standard loss functions without class frequency consideration, then training converges quickly on dominant classes, but recognition accuracy for less frequent classes deteriorates
Solution Approach 1:
The patent modifies the loss function parameters by incorporating class frequency distribution into the weighting scheme. This parameter change allows the optimization process to balance convergence speed with improved accuracy for rare classes, as the weighted loss function guides the learning process to allocate more attention to underrepresented gestures during training
Solution Approach 2:
The patent implements feedback mechanisms by using the class frequency distribution as a guiding signal during training. The loss function continuously adjusts based on the observed class frequencies, providing feedback to the optimization process that helps balance the treatment of different gesture classes and prevents dominance by frequent gestures
Data Source
AI summary
A method for automatic gesture recognition in which, by a machine learner data of a respective gesture of a variety of gestures executed by a user and captured by at least one sensor gestures of a class are assigned to a variety of predetermined classes, and in which the machine learner is trained with a training data set, which is divided into predetermined data segments. The respective data segments are data segments of the training data set are assigned to the variety of predetermined classes by means of the machine learner. A respective contribution, with the respective assignment processes for the assignment of respective data segments are incorporated into the automatic gesture recognition of a particular class.