galmetzer / dipole-normal-prop Goto Github PK

Hi, Gal.

I want to re-implement the results of Fig. 16 in the main paper.

It is tough for me to tune the suitable parameters since the input kitten from the VIPSS dataset is very sparse (500 pts).

I directly used the parameters of the demo/fandisk.sh with Possison Rec, but it seems did not work very well...

Would you like to share the parameters for the sparse kitten? Thanks a lot!

Looking forward to your repleyment.

Hello,

First thank you for this very usefull algorithm.

I am trying to use it on point clouds generated from segmentated objects of a noisy dataset (from cryo-tomography of biological samples) but I am not sure how to set the parameters.

On small patches with one layer of points it works well.
(object 1)

(object 2)

But when there are two layers of points, the normals are not oriented in opposite directions from one layer to the other as you can see here with a small connection between the two layers which i want :

(top view)

(side view)

here with no connection between the two layers :

(full view)

(Zoom 1)

(Zoom 2)

The point clouds contained between 10 000 and 20 000 points (it has been subsampled from 100 000 points).

The parameters used are the following :

python -u $BASE_PATH/orient_pointcloud.py
--models $BASE_PATH/pre_trained/hands2.pt
$BASE_PATH/pre_trained/hands.pt
$BASE_PATH/pre_trained/manmade.pt
--iters 10
--propagation_iters 5
--number_parts 30
--minimum_points_per_patch 100
--weighted_prop
--estimate_normals
--diffuse
--curvature_threshold 0.9 \

To obtain this I had to use orient_large.py. With orient orient_pointcloud.py there are small inconsistent patches. Also I had to use curvature_threshold 0.9 for the same reason.

Do you know what parameters should I add/delete/tune to make the two layers consistent ?

Thank you for your help !

Julien Maufront

Hello,Gal
my GPU memory is 3GB that can't train the large point clouds(like lion),could you tell me which parameter is about setting the batch_size?