dataglove_manus-prime-x_augmented-handshapes

IMPORTANT NOTE: The datasets have user-dependency issues and should not be used for classification without this limitation in mind. Furthermore, augmentation should be done "live" instead of using these datasets (see gesture-classifier repository).

This repository contains the augmented data based on the available data recorded with the Manus Prime X Haptic VR data glove.

File Structure

• augmentation_method
  
  • n_fold
    
    • results
      
      • 29_gestures
        
      • 62_gestures
        
    • static
      
      • ...
        
    • info.txt
      

For each method there is a folder with the same name. They contain datasets with different sizes and the corresponding recognition results (either for 62 or 29 gesture labels). An info.txt file includes information about the parameters that led to the results. The static folder contains the data.

Data Format

The augmented data has the same format as the non-augmented data for the MANUS data glove, see Manus Prime X Data Glove Handshapes. The synthethically generated data samples are simply appended at the end of the available files for each user and static gesture.

Methods

Note: Not all datasets generated by the proposed data augmentation methods were able to improve the classifiers performance. Some techniques even decreased the performance. In short, Feature Shuffling, Range Based and Range Interpolation, as well as their combinations, proved to be successful. The rest did not.

Random Noise

Adds artificial noise to the available data.

Scaling

Scales the training data either up or down by multiplying it by a random scalar.

Range Based

Generates synthetic training data by randomly sampling within defined limits for each joint/feature. The limits are chosen by calculating the minimum and maximum for each feature value and sampling new values within the range [min, max].

Mean/Median Based

Calculates the mean/median feature values for each label and generates similar, synthetic training data.

Feature Shuffling

Generates new data samples by shuffling/permuting the feature values between all data samples available for a specific gesture.

Range Interpolation

Generates new data samples by linearly interpolating between the minimum and maximum feature value for each joint.

Inter User Interpolation

Same as Range Interpolation, but only interpolates between the minimum and maximum feature values for each joint and a user.

Dynamic Data

Does not generate new data, but simply extracts additional data samples from the raw dynamic data recorded with the data glove.

Sensor Dropout

Randomly drops one of the sensor/feature values with a certain probability, i.e. sets it to zero.

Combinations

The best individual methods are combined into one larger dataset.

Pipeline Combinations

The best individual methods are combined into one larger dataset, however the methods are pipelined so that the already augmented data is augmented again.