Figure 6.

Effect on shoulder girdle stabilization exercise inclusion/exclusion from training data for OOD detection of the SPARS9x dataset on: (a) FCN Softmax activation; and (b) normalized KNN nearest neighbor distance. This figure highlights the resiliency to changes in training class inclusion/exclusion of the KNN algorithm compared to deep learning algorithms for detecting OOD data with this dataset. Note that, even though the mean Softmax activation decreases for the OOD data and increases for the in-distribution classes with shoulder girdle stabilization removed, there is still a slight decrease in AUROC (see Softmax Threshold of Figure 7b).