About the speed

[seed93]

Have you tested the speed? It run over 30 seconds for a 10k vertices mesh, which is abnormal compared with open3d. I checked it was compiled with openmp. Thank you. It is a great work.

[f1shel]

Oooops, I have encountered similar issues actually. This package seems much slower than other UV unwrapping libraries. I mainly followed the logic in UVAtlas.cpp, eliminating all optional branches. Could you please run the official Microsoft UVAtlasTool on the same mesh and let me know how long it takes?

[seed93]

Oooops, I have encountered similar issues actually. This package seems much slower than other UV unwrapping libraries. I mainly followed the logic in UVAtlas.cpp, eliminating all optional branches. Could you please run the official Microsoft UVAtlasTool on the same mesh and let me know how long it takes?

I didn't try the official tool, but it is much faster as far as I know. If it was me to debug, I would profile which part is most time consuming. I guess it only uses one cpu core to compute while openmp is enabled. Also maybe there are some optimization methods not suitable for linux.

[seed93]

https://github.com/isl-org/Open3D/blob/main/cpp/open3d/t/geometry/kernel/UVUnwrapping.cpp in open3d, it use multithread above uvatlas. Maybe it is the key point.

[f1shel]

Oooops, I have encountered similar issues actually. This package seems much slower than other UV unwrapping libraries. I mainly followed the logic in UVAtlas.cpp, eliminating all optional branches. Could you please run the official Microsoft UVAtlasTool on the same mesh and let me know how long it takes?

I didn't try the official tool, but it is much faster as far as I know. If it was me to debug, I would profile which part is most time consuming. I guess it only uses one cpu core to compute while openmp is enabled. Also maybe there are some optimization methods not suitable for linux.

Thank you for the feedback :] Sorry that I'm a bit busy with my paper right now. I estimate I will have time to look at this bug in a few weeks. I'll look into it right away as soon as I have some free time.

[seed93]

It seems we need a pcapartition and then process parallely. https://github.com/isl-org/Open3D/blob/main/cpp/open3d/t/geometry/kernel/PCAPartition.cpp

Will you improve this?

[f1shel]

It seems we need a pcapartition and then process parallely. https://github.com/isl-org/Open3D/blob/main/cpp/open3d/t/geometry/kernel/PCAPartition.cpp

Will you improve this?

I'll take a look. I have also noticed that the parameter parallel_partitions in open3d.cpu.pybind.t.geometry.TriangleMesh.compute_uvatlas significantly affects the speed of UV mapping. However, it's strange that I can't seem to get the correct UV partitions when using it. Download bunny.obj here.

import open3d as o3d
import numpy as np
import trimesh
import time

mesh = trimesh.load('bunny.obj', process=False, force='mesh')

o3d_mesh = o3d.t.geometry.TriangleMesh()
o3d_mesh.vertex.positions = o3d.core.Tensor(mesh.vertices)
o3d_mesh.triangle.indices = o3d.core.Tensor(mesh.faces)

t1 = time.time()
o3d_mesh.compute_uvatlas(size=1024, gutter=10, parallel_partitions=4)
print("Time:", time.time() - t1)

# since in open3d uv is defined on triangles, we need to rearrange vertices
new_v = mesh.vertices[mesh.faces.reshape(-1)]
new_f = np.arange(len(new_v)).reshape(-1,3)
new_uv = o3d_mesh.triangle.texture_uvs.numpy().reshape(-1,2)

mesh = trimesh.Trimesh(
    vertices=new_v,
    faces=new_f,
    process=False, # !important
)
mesh.visual = trimesh.visual.TextureVisuals(
    uv=new_uv.astype(np.float32), 
)
ret = mesh.export('bunny.uv.obj')

[f1shel]

It seems we need a pcapartition and then process parallely. https://github.com/isl-org/Open3D/blob/main/cpp/open3d/t/geometry/kernel/PCAPartition.cpp Will you improve this?

I'll take a look. I have also noticed that the parameter parallel_partitions in open3d.cpu.pybind.t.geometry.TriangleMesh.compute_uvatlas significantly affects the speed of UV mapping. However, it's strange that I can't seem to get the correct UV partitions when using it. Download bunny.obj here.

import open3d as o3d
import numpy as np
import trimesh
import time

mesh = trimesh.load('bunny.obj', process=False, force='mesh')

o3d_mesh = o3d.t.geometry.TriangleMesh()
o3d_mesh.vertex.positions = o3d.core.Tensor(mesh.vertices)
o3d_mesh.triangle.indices = o3d.core.Tensor(mesh.faces)

t1 = time.time()
o3d_mesh.compute_uvatlas(size=1024, gutter=10, parallel_partitions=4)
print("Time:", time.time() - t1)

# since in open3d uv is defined on triangles, we need to rearrange vertices
new_v = mesh.vertices[mesh.faces.reshape(-1)]
new_f = np.arange(len(new_v)).reshape(-1,3)
new_uv = o3d_mesh.triangle.texture_uvs.numpy().reshape(-1,2)

mesh = trimesh.Trimesh(
    vertices=new_v,
    faces=new_f,
)
mesh.visual = trimesh.visual.TextureVisuals(
    uv=new_uv.astype(np.float32), 
)
ret = mesh.export('bunny.uv.obj')

I've identified the bug: it's important to remember to pass process=False when initializing a trimesh.Trimesh. Now, the issue is that I wasn't aware Open3D already has a similar UVAtlas tool when I initially wrapped this package. After some testing, I found that Open3D performs better, although it requires several rounds of conversion. Given this, is it still necessary to update this repository?