Partiotioning .ply file.

[kcabhishek]

I am trying to do partition on a .ply file, will I be able to do that with this? If so, how?

[loicland]

You need just a few modifications:

• partition/partition.py line 42, if you don't want to use a train/test split
• line 136-138: comment the call to read_las if your cloud is a ply
• you might need to alter the function read_ply in provider.py to read your data format.

Then run:

python partition/partition.py --dataset custom_dataset --ROOT_PATH data_location --voxel_width 0.05 --reg_strength 0.5

values for voxel_width should be adapted so that your cloud have more or less uniform sampling precision, and reg_strength depending on the desired granularity of the partition

Then use the function visualize with --output_type ip to see the results.

[kcabhishek]

I am running with this.

python partition/partition.py --dataset custom_dataset --ROOT_PATH partition/data/ --voxal_width 0.05 --reg_strength 0.5

Let me explain what I am trying to achieve here so that I won't waste your precious time asking something that this code isn't supposed to do. So, I have a 3D model of a office and my goal is to partition or segment the objects in that office. Like chairs, tables, etc. Am I in the right place? Thank you for your help though. I really appreciate it.

[loicland]

It's voxel not voxal.

Am I in the right place?

Sure, provided that your office is similar to the ones in S3DIS, or that you have enough training data to train a model from scratch. This repo won't give you instance/object segmentation however, just the class of each points.

[kcabhishek]

What kind of data do I need to train my model from scratch? Also what do you mean by just the class of each points?

[loicland]

You need at least a dozen scenes with points annotated by semantic class.

The output of SPG is a class for each points. If you want object detection or instance segmentation you should use a other network. Checkout torch-points3d for example, they have most tasks implemented.