ST-GCN Skeleton Action Detection Dataset Denoising and Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The ST-GCN model for skeleton-based action recognition has limitations in accuracy, particularly in recognizing actions like fighting, where it may misclassify actions such as hitting, wielding a knife, and pushing.
Innovation Solution
To improve the accuracy of the ST-GCN model without compromising speed, various pre-processing data augmentation steps are applied, including 2D rotation, 2D shear, scaling, horizontal flips, and the addition of Gaussian noise, along with denoising processes for training and input data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If the ST-GCN model is used for fast and small processing, then speed is improved, but accuracy deteriorates
Solution Approach 1:
The patent applies data augmentation techniques (2D rotation, 2D shear, scaling, horizontal flips, Gaussian noise addition) and denoising processes to the training dataset before model training. This preliminary processing of data improves the model's accuracy without requiring changes to the model architecture, thus maintaining the speed advantage while enhancing recognition performance.
2Measurement precision
If data augmentation techniques are applied, then accuracy is improved, but dataset complexity increases
Solution Approach 1:
The patent transforms the original skeleton joint data by applying various parameter changes including 2D rotation, 2D shear, scaling transformations, horizontal flips, and Gaussian noise addition. These parameter modifications create diverse training samples from existing data, improving model accuracy while using computationally efficient transformations that don't significantly increase processing complexity.
Data Source
AI summary
Disclosed here are various techniques for improving the testing and training of datasets comprising sequences of skeletal representations performing various actions. The dataset can be denoised by applying various techniques to determine noisy frames within each sequence and eliminating the sequences from the dataset when the number of noisy frames in the sequence is too large. In addition, the dataset may be augmented by various data augmentation techniques to manipulate the skeletal representations, after denoising.