Point clouds to mesh MeshLAB

Procedure with Meshlab free program

http://www.meshlab.net/

The “Filter->Sampling->Poisson Disk Sampling
 

Make sure you check the “Base Mesh Subsampling” box.

Filters -> Point Set -> Compute Normals for point set”

The “Filters -> Point Set-> Surface Reconstruction: Poisson”

 

  Detailed explanation

Occasionally you will need to sub-sample your point-cloud data to
make it easier to work with. This does inevitably reduce the resolution of the
data but if proper techniques are used you can maintain a high level of
fidelity in the point cloud data.
*** Especially in noisy scan’s from the Kinect

We will want to
recreate a surface, which through trial and error (at least with objects that contain
  a lot of curves or contours) the Poisson disk method obtains the
best results.
The
“Filter->Sampling->Poisson Disk Sampling”

Make sure you check the “Base Mesh Subsampling” box.


The algotrithim
 it was designed to create circular window over the point cloud and
calculate those points that are statistically “random” according to a
Poisson distribution.

Like previously mentioned the exact parameters used in your
process are TOTALLY APPLICATION DEPENDENT. Meaning that what worked well
with a point cloud of a million points for the interior of a room, may not work
with a million points of a human face.

The image below the point cloud captured from the Microsoft Kinect
(of a human chest – side view) and it has points that are not apart of the
actual object we want to creat a 3D model of. So to avoid have spikes or
deformities in our data we should apply a few methods in eliminating them when
possible. False points to be removed from point set data

While there are many different ways to deal with these rouge points we can once again apply the Poisson distribution, which seems to have the best results in the automated
filters offered by MeshLab. 

Much like the
filtering of noise in LiDAR data the Poisson takes the entire area of
interest(the radius of the window size we specify in this case) and looks at
the corresponding distribution of points in 3D space. When points are
determined to be statistically random following the number of iterations you
specify the alogritim will remove that point from the recreation of the
surface. 

Even though the Poisson does an excellent job there are still cases where manually cleaning these points from the data is required. (Meaning select it and delete it)

It is also important to note that since the Poisson is a
stochastic process no two subsamples will be exactly the same even if the exact
same parameters are used. So save your data often!!

We will now have to calculate the normals on the sub-sample
we just created so MeshLab knows which side of the point is facing “out” and
which is “in”.

For the point set:

“Filters -> Point Set -> Compute Normals for point set”

At this point you will need to choose one of the surface
reconstruction algorithms that MeshLab offers.

The “Filters ->
Point Set-> Surface Reconstruction: Poisson”

*** Note: This could get time consuming and at least in my experience crashes
when the data  is huge(“huge” is a scientific word for bigger than normal)

As mentioned before in
the subsampling discussion a few tabs ago you can also use the “Marching Cubes
(APSS)” which has pretty good results on data with few contours.

For you inquisitive folks who need to know more about each
of these processes for surface reconstruction please check out these two
links: Marching Cubes or the Poisson