Table 3.

Overview of all final benchmark results with average classification accuracy and average F1-score (in brackets) for the chosen configuration of each approach. The last column shows the configuration’s hyperparameters, as well as the $\gamma$ trained for outlier detection.

Approach	KNN		SVM		Outlier Detect	Configuration
	Pocket	Hand	Pocket	Hand
Analytical Transform	$92.4$% ($91.6$%)	$76.4$% ($89.1$%)	$96.0$% ($95.2$%)	$84.1$% ($85.3$%)	$95.5$% $(\gamma =0.005)$	PCA 5-dim, $w_{\mathrm{seq}}=2\mathrm{s}$, raw acceleration
Codebook	$95.5$% ($98.1$%)	$68.1$% ($88.3$%)	$98.4$% ($99.1$%)	$76.6$% ($91.5$%)	$86.0$% $(\gamma =15.0)$	$\sigma =0.25$, $\left|\mathcal{C}\right|=32$, $w_{\mathrm{sub}}=32,w_{\mathrm{seq}}\approx 2\mathrm{s}$, raw acceleration
Statistical Features	$95.5$% ($98.3$%)	$72.2$% ($87.5$%)	$95.4$% ($98.3$%)	$69.9$% ($87.6$%)	$92.4$% $(\gamma =4.55)$	LDA 5-dim, $w_{\mathrm{seq}}=2.5\mathrm{s}$, $n_{\mathrm{mel}}=10$, $\psi =1.6$, $\left|\mathit{q}\right|=5$, two-channel acceleration