search

compas-dev / compas

Core packages of the COMPAS framework.
https://compas.dev/compas/
MIT License
313 stars 108 forks source link

Calculation of geodesic distances starting from many vertices #538

Closed ioannaMitropoulou closed 4 years ago

ioannaMitropoulou commented 4 years ago

Describe the bug I would like to report a problem I have with the heat geodesics calculation (compas.datastructures.mesh_geodesic_distances_numpy). When I calculate geodesic distances with one vertex as a start point, the result seems to be correct (though I haven't verified it). However, when I calculate with multiple start points, then the result is definitely wrong. (see screenshots below)

To Reproduce I have created this simple example that illustrates the problem. You can download it here: https://www.dropbox.com/sh/qg10kywmiugh52c/AADV7J0VQkB9hDmgoZ-pEZeza?dl=0 Run the code to see the distance gradient that is produced starting from multiple vertices.

Expected behavior The correct gradient starting from all these vertices should be parallel to the boundary line. As a comparison, I also attach images of the geodesic distances (again using the same heat algorithm), calculated in libigl, with the exact same inputs.

Screenshots

Compas

Geodesic distances starting from one vertex. These seem to be correct. (starting vertex in blue)

3b

Geodesic distances starting multiple vertices. These are wrong. (starting vertices in blue)

3a

Libigl Geodesic distances starting from one vertex. (starting vertex in red)

3d

Geodesic distances starting multiple vertices. (starting vertices in red)

3c

Desktop (please complete the following information):

tomvanmele commented 4 years ago

with a value of 1000 for parameter m, the result is (almost) the same as the image from igl...

ioannaMitropoulou commented 4 years ago

Many thanks for the reply! Indeed you are right. I had not tried using such a high value for m, because in the heat geodesics paper a time step with m=1 is recommended. It says that higher values would create a smooth approximation of the geodesic distances. But maybe this m=1000 just accounts for the different units? I will test it with m=1000 on more complex shapes and will let you know if I have more problems.

tomvanmele commented 4 years ago

in any case the exact method and the heat method of libigl are available through compas_libigl, which by the end of this week should be conda installable...

ioannaMitropoulou commented 4 years ago

Great, thanks for the info! On the same example, with m = 1000 and a some better coloring, I still see some problems in the resulting distances.

3e