Poor Output from PLY

[AmanKishore]

I'm getting really poor output from the PLY:

Am I using marching_cubes_mesh incorrectly?

mesh = marching_cubes_mesh(
    pc=pc,
    model=model,
    batch_size=4096,
    grid_size=256, # increase to 128 for resolution used in evals
    progress=True,
)

[giddyyupp]

I have the same problem.

[fractaltr33]

Now I have an error

AttributeError                            Traceback (most recent call last)
[<ipython-input-12-9f81ccec745d>](https://localhost:8080/#) in <module>
      1 # Produce a mesh (with vertex colors)
----> 2 mesh = marching_cubes_mesh(
      3     pc=pc,
      4     model=model,
      5     batch_size=4096,

[/content/point-e/point_e/util/pc_to_mesh.py](https://localhost:8080/#) in marching_cubes_mesh(pc, model, batch_size, grid_size, side_length, fill_vertex_channels, progress)
     68         volume_np -= np.mean(volume_np)
     69 
---> 70     verts, faces, normals, _ = skimage.measure.marching_cubes(
     71         volume=volume_np,
     72         level=0,

AttributeError: module 'skimage' has no attribute 'measure'

[fractaltr33]

And I have a question, how can I save my .npz file?

[AmanKishore]

@fractaltr33 Check out #8 for saving .npz file. You probably need to do pip install --upgrade scikit-image

[fractaltr33]

@AmanKishore I'm far from programming, how should I use save() in colab?

[AmanKishore]

Add this line at the end of the file pc.save('pointcloud.npz')

[unixpickle]

The quality issue in the initial post is likely not a bug, but an actual case where the point cloud -> mesh model doesn't do a great job. There are some tricks that might improve quality, like adding some bias to the SDF predictions to thicken the object, or applying smoothing to the points in the point cloud, but there will always be some cases like this.

In the future, the meshing quality could be improved by using other, more advanced existing techniques like SAP. The method we use based on SDF regression can suffer with thin parts or noisy point clouds, as shown in the paper. This is probably a bigger problem with the small text2pc model, since the model is less large and therefore less capable of producing clean pointclouds.