Table 1.

The classification rates of the global separators introduced in Section 2.3—the linear separators ${\mathbf{w}}_{\mathrm{UV}}$ (UV-only), ${\mathbf{w}}_{\mathrm{G}}$ (green-only), ${\mathbf{w}}_{\mathrm{con}}$ (contrast: 1:1), ${\mathbf{w}}_{\mathrm{F}}$ (contrast: Fisher discriminant) and the binary masks (non-linear UV/G separator)—applied to the datasets shown in Figure 9 and Figure 10. The separation methods were trained and tested on two different samples (each containing ${10}^5$ elements of the sky and ground class) drawn from the specified databases. An evaluation using single HDR image pairs (e.g., as captured by a mobile robot) as test data instead can be found in Figure 11. Only for sand (highlighted), the classification rates show a noticeably increased performance for the UV/G contrast (Fisher discriminant) compared to UV-only separation. In all other cases, both methods show a similar performance, both slightly worse compared to the best possible performance of the binary masks. Classification rates obtained by applying the learned global separation techniques to the input HDR images (UV and green) directly can be found in Figure 11.

Global	${\mathit{X}}_{8–19}$
	UV	Green	Contrast	Fisher	Mask
Stones	$84\%$	$78\%$	$70\%$	$84\%$	$88\%$
Sand	$\mathbf{88}\mathbf{\%}$	$\mathbf{60}\mathbf{\%}$	$\mathbf{89}\mathbf{\%}$	$\mathbf{94}\mathbf{\%}$	$\mathbf{95}\mathbf{\%}$
Earth	$94\%$	$88\%$	$78\%$	$94\%$	$95\%$
Forest/Suburban	$95\%$	$92\%$	$80\%$	$96\%$	$97\%$
All	$88\%$	$79\%$	$79\%$	$89\%$	$92\%$
Global	${\mathit{X}}_{7–20}$
	UV	Green	Contrast	Fisher	Mask
Stones	$79\%$	$75\%$	$67\%$	$79\%$	$84\%$
Sand	$\mathbf{83}\mathbf{\%}$	$\mathbf{60}\mathbf{\%}$	$\mathbf{86}\mathbf{\%}$	$\mathbf{89}\mathbf{\%}$	$\mathbf{91}\mathbf{\%}$
Earth	$89\%$	$84\%$	$73\%$	$88\%$	$90\%$
Forest/Suburban	$91\%$	$89\%$	$77\%$	$92\%$	$93\%$
All	$83\%$	$75\%$	$75\%$	$84\%$	$87\%$