Tips for creating PointMeshes of faces (and other body parts)

To obtain the best results, in addition to the regular requirements, modeling faces with PhotoModeler Scanner requires the following:

1) Well oriented synchronized cameras.  The project should use a calibrated camera, sub-pixel targets and have a largest residual of less than one. Due to the already weak texture of a face any subtle movements will affect the DSM scan. You may get an approximate result using one camera (taking one stereo pair) if the subject is very very still, but using two or more synchronized cameras will certainly improve the result.  If you find you are not getting good results, or the amount of noise it too high, then changing to synchronized cameras may give you your largest improvement.  One lower cost method of synchronizing cameras with built in IR sensor is to use a single IR remote to fire all cameras at once.

2) On the Dense Surface dialog: lowering the texture type to 1 to reduce the number of dropped points and then increasing the matching region radius if a lot of points are still dropped or to reduce noise.

3) Starting with a reasonable approximate surface so the search depth range can be lowered.  That is the project should contain points that give a rough shape to the face and not just target points used for orientation.  For example adding a manually referenced point on the tip of the nose.

4) Multiple pairs from multiple synchronized cameras will be needed to increase coverage to surround more of the face and capture recessed or raised areas blocked by occlusions in other pairs. A combination of vertical and horizontal stereo pairs will give the most coverage. If multiple pairs are used with a low amount of overlap it may be best to use the ‘Merge only’ option of the register and merge meshing step.
5) Hair, glasses or other secular objects can cause additional noise and may not be captured well.

6) If the goal is a textured low poly count triangulated mesh using a larger sampling rate will produce better results than a small sampling rate creating a high density point cloud and then decimating the triangulated mesh.

7) If ripples are seen in the resulting triangulation the sampling rate may be too high for the image data.  Either lower the sampling rate (to get a more sparse point cloud) or increase the number of pixels used to image the face.

8) Due to the low texture quality of faces often the mesh will require some post processing clean-up using the point mesh edit tools.

Click here to view the biological examples on the samples page.