Results of the dual relaxation are shown in Fig. 2.
One scan of a femur was acquired axially at a resolution of 
. The other scan (of
the same person) was acquired sagittally (one year later) at a
resolution of 
. The femur was segmented manually from both
images
. Fig. 2a,b
shows the results of running Marching Cubes
[4], individual SurfaceNets [2], and
Dual SurfaceNets on the images. No decimation was performed on any of
the models. Notice the terracing artifacts in the models generated
with Marching Cubes and individual SurfaceNets along the direction
that the scans were acquired. The model generated using Dual
SurfaceNets on both scans preserves the fine details in the original
scans well but does not contain the terraces.
In the second example, we consider building a model from extremely low
resolution scans. Fig. 2c shows results of model
generation from subsampled versions of the original segmentations.
The axial and sagittal scans were subsampled by a factor of 4 to
resolutions of 
and 
respectively. The model generated using Dual SurfaceNets at the low
resolution contains slightly less detail than the high resolution
version, but it is remarkably smooth and free of terracing artifacts,
while remaining faithful to the original segmentations.
The surface models can be visually verified by superimposing the
relaxed net on the image data. Fig. 2d shows the
input segmentations to the Dual SurfaceNet algorithm and the final
result of the net.
Despite the
blockiness evident in all the input segmentations, the final models
are very smooth and capture the details of the femur.