Figure 1.5 : The GroEL molecular structure and an engine block. 1.5(a) shows
the raw density map, and 1.5(b) is the result of its segmentation into 14 subunits
using our tool. 1.5(c) shows the raw engine data and 1.5(d) shows the result of the
segmentation with transparency.
Figure 1.5 shows raw volumes before and after segmentation.
Many researchers have studied the problem of 3D segmentation in different do-
mains. In particular, automatic methods for segmenting volume data have received
much attention. Automatic techniques usually involve quantifying low-level features
to identify regions of interest in each specific data set. Such features include similarity
in values between voxels, proximity of voxels, and values of high-order derivatives of
the implicit function. Various techniques have been explored on this front including
thresholding, region growing, and water-shed [29]. However, automatic methods suffer
from an excess of parameter manipulation. The user often has to repeat the process
of inputting parameters and executing the algorithm to perform segmentation. More-
over, each automatic segmentation technique is tailored towards a specific problem.
Transferring a technique across problem domains often results in poor segmentations
because segmentations often require semantic interpretations and different domains
have different semantics attach to their segmentation. Emulating human semantics in
one domain is already a difficult task, and trying to emulate all the possible semantics
that could arise in all domains is a far more daunting problem.
In contrast to a fully automatic approach, other works focus on semi-automatic