Fig. 5.

Parametric surfaces reconstructed with the different RBF. (a) $\mathrm{\Delta }c{\mathrm{HbO}}_2$ and (b) $\mathrm{\Delta }c\mathrm{HbR}$. Each column corresponds to a different RBF. For each subsampling scenario, rows correspond to approximated interpolations. Being piecewise interpolation, training with fewer samples yields smoother surfaces without severely penalizing the error. The orientation of the surfaces is arbitrary as dictated by the SVD projection, and results from the ill-posed problem of dimensionality reduction in which flipped or rotated solutions are equivalent. A trend is apparent in most cases; along with the main component for $\mathrm{\Delta }c{\mathrm{HbO}}_2$ and tilted over the two main components for $\mathrm{\Delta }c\mathrm{HbR}$.