pointcloudmatcher

Code for the paper: "Cluster-Based Autoencoders for Volumetric Point Clouds" arXiv:2211.01009.

Installation

Create a python virtual environment and pytorch (preferably version 1.11) and pytorch3d (tested with version 0.6.2).

Then one also needs to install the following packages manually since we used non-release versions for them.

• Geomloss It is best to install the version from pull#49 since it contains some performance optimizations for evaluation.

git clone https://github.com/jeanfeydy/geomloss && cd geomloss
git remote add 2origin https://github.com/martenlienen/geomloss
git pull 2origin
git checkout e6ff657d4a063bfb08351c08fc3e11eb55d0c319
pip install -e .

• kmeans-pytorch Install the latest version via git i.e. pip install 'git+https://github.com/subhadarship/kmeans_pytorch' or to get the exact version

git clone https://github.com/subhadarship/kmeans_pytorch && cd kmeans_pytorch
git checkout f7f36bd1cb4e3a761d73d584866d0a9c6b4d2805
pip install -e .

• k-means-constrained

pip install 'numpy<1.23'
pip  install  --no-build-isolation k-means-constrained==0.6.0 --no-binary k-means-constrained

Then install the rest with

pip install -e .

Usage

import numpy as np
import sfb

# load pretrained model
model = sfb.load_network("weights/foldingnet.ckpt")

# load pointcloud
x = ...

# define interpolation factor
alpha = 0.2

design_pc = sfb.density.normalize_design_pc(np.loadtxt("pointclouds/stripe.xyz"))

z = sfb.apply_design(model, x, design_pc, model.config["num_points"], alpha)

Training

In order to train a model with your own dataset one first needs to apply the design point clouds (optional) and then cluster the pointclouds (unless they are already of the desired size). We provide a function to preprocess the data preparedata which works for xyz point clouds. If on has xyz point clouds in the folder /data/ and a design point cloud /example/design.xyz, then use

sfb.preparedata(["/data/"], ["/example/design.xyz"], "/output/", 2048)

This will create folders /output/applied/design with design.xyz applied to the point clouds and /output/combined_split with the individual clusters of all point clouds (in binary format).

Then one can train a model

python train.py --num_points=2048 --emd --blosc --datroot=/output/combined_split --factors_file=weights/factors.json

See python train.py --help for additional options.