About 10000 points

[waltstephen]

When you used sketchfab to build the dataset, the total number of point clouds was limited to 10,000. What transformation was used to achieve this?

[seanzhuh]

I have the same question, how to sample 10,000 points from mesh surface, is it fps?

[seanzhuh]

I have the same question, how to sample 10,000 points from mesh surface, is it fps?

And do you randomly sample points during training to make the model robust to the number of input points? Since not all objects have the same number of points in general.

[seanzhuh]

I have the same question, how to sample 10,000 points from mesh surface, is it fps?

And do you randomly sample points during training to make the model robust to the number of input points? Since not all objects have the same number of points in general.

While I just found that the authors mask out 50% point tokens during training by enabling the PatchDropout before feeding point tokens to point transformer, is this equivalent to randomly drop points ?

[waltstephen]

In fact, I finally chose Poisson sampling to sample 10,000 points for processing, and the effect was good.

[seanzhuh]

In fact, I finally chose Poisson sampling to sample 10,000 points for processing, and the effect was good.

Could you please share more details?

[waltstephen]

pcd = mesh.sample_points_poisson_disk(number_of_points=num_samples, init_factor=5)     points = np.asarray(pcd.points)
Then downsample to 10000

[seanzhuh]

thanks a lot，where is this mesh imported?