Graphics Reference
In-Depth Information
where
t
denotes the texel on the surface. A primary advantage of the TensorTex-
tures model is that it works well even for a sparsely sampled BTF data set.
The TensorTextures paper contains a comparison of the accuracy of BTF ap-
proximation using PCA with that using the
M
-mode SVD, in terms of both numer-
ical accuracy and rendered image quality. For the same number of basis vectors,
the TensorTextures method is generally better in both respects. Furthermore, the
authors contend that even when the TensorTextures method incurs a larger numer-
ical error, the rendered images tend to have better visual quality.
Two papers were presented at SIGGRAPH 2006 that considered the repre-
sentation of time-varying reflectance character, i.e., the temporal extension of a
BTF [Gu et al. 06, Wang et al. 06]. Matrix-based methods are not well suited
to this problem, as they are inherently two-dimensional. In contrast, the Tensor-
Textures method can be extended to include time variation by simply adding an-
other dimension (mode) to the data tensor. A drawback of TensorTextures and the
matrix-based methods alike is the difficulty in representing “high frequency” re-
flectance characteristics, such as highly specular reflection. Nonetheless, Tensor-
Textures seems a promising approach for use in future computer graphics appli-
cations.
