Handed Activity Recognition With Horizontal Flipping for Dataset Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based systems struggle to accurately distinguish between left- and right-handed activities due to the imbalance in the population distribution of left- and right-handed individuals, leading to increased complexity and cost, and require separate classes for each, which is challenging to balance in training datasets.
Innovation Solution
An activity recognition system that recognizes activities of one handedness and infers them twice with horizontal flipping to identify both left- and right-handed activities, combined with a training method that flips all training instances to balance the dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the system creates separate classes for left- and right-handed activities, then the accuracy in distinguishing handed activities is improved, but the device complexity and training cost increase
Solution Approach 1:
The patent applies asymmetry by creating a single asymmetric activity class that can represent both left-handed and right-handed variations. Instead of creating separate symmetric classes for each handedness, the system uses one class with asymmetric handling through spatial transformation (horizontal flipping) during training, allowing the model to distinguish handedness without doubling the number of classes.
Solution Approach 2:
The patent implements universality by designing a single activity class that serves multiple functions - it can represent both left-handed and right-handed activities through the use of spatial flipping during training. This multi-functional class reduces the need for separate dedicated classes while maintaining the ability to accurately distinguish between different-handed activities.
2Measurement precision
If the training dataset includes equal numbers of left- and right-handed activity examples, then the classification accuracy for both handednesses is improved, but the data collection cost and time increase
Solution Approach 1:
The patent applies copying by creating synthetic training data through horizontal flipping of existing video clips. Instead of collecting separate left-handed and right-handed activity data from multiple sources, the system takes a single handedness example and generates its mirror image as training data, effectively duplicating the training set without requiring additional time-consuming data collection efforts.
Solution Approach 2:
The patent implements preliminary action by pre-processing the training data through horizontal flipping before training begins. This preliminary transformation ensures that the training dataset automatically achieves balanced representation of both handednesses without requiring post-collection balancing, saving time in the data preparation phase.
3Device complexity
If the system uses a single activity class for both left- and right-handed activities, then the device complexity is reduced, but the ability to distinguish between handednesses deteriorates
Solution Approach 1:
The patent resolves this contradiction by introducing asymmetric handling within the single class through spatial flipping operations. The single activity class maintains asymmetric properties by flipping input images horizontally during training, enabling the model to learn handedness distinctions without creating multiple symmetric classes.
Solution Approach 2:
The patent applies inversion by using horizontal flipping as a transformation technique. Instead of creating separate classes for different handednesses, the system inverts the input space by flipping images, allowing a single class to capture both handedness variations through this spatial inversion technique.
Data Source
AI summary
This disclosure describes an activity recognition system for asymmetric (e.g., left- and right-handed) activities that leverages the symmetry intrinsic to most human and animal bodies. Specifically, described is 1) a human activity recognition system that only recognizes handed activities but is inferenced twice, once with input flipped, to identify both left- and right-handed activities and 2) a training method for learning-based implementations of the aforementioned system that flips all training instances (and associated labels) to appear left-handed and in doing so, balances the training dataset between left- and right-handed activities.