As motivated in Section 5, information from different views will
be naturally misaligned since the proxy geometry is an approximation
of the hair volume and the appearance of hair may differ
substantially in different views. Consequently we merge lowfrequency
color information by averaging the contributions of the
individual views to avoid color seams (Section 5.1), but we must
combine high-frequency components using the single-best view to
avoid blurring (Section 5.2). As a result, the high-frequency color
information on the mesh will contain seams, and this can adversely
impact orientation tensor computation if performed directly on the
mesh. This is demonstrated in Figure 9 on a synthetic mesh patch.
The high-frequency components are sampled from three different
views (color-coded for better visualization) using the single-best
view per vertex (Figure 9 (a)). The seams are clearly visible due to
the change in orientation of the features. If we compute the orientation
tensor on the mesh, it will inevitably follow such seams and so
will do the directional smoothing (Figure 9 (b) and (f)). As a consequence,
we will compute the orientation tensors from the images
and transfer them to the mesh vertices.