POint cloud LAtent Registration

Getting started

POLAR is a method to simultaneously register numerous highly degraded point clouds corresponding to views of the same unknown reference object.

Installation

Warning

PyTorch3D is required by POLAR. Since its installation is quite specific, you need to install it manually.

If you want to user POLAR with the provided pretrained autoencoder, run

pip install polaregistration

If you want to retrain the autoencoder on your own, run

pip install polaregistration[train]

If you want to use the interactive visualization function, run

pip install polaregistration[vis]

Finally, to install everything, run

pip install polaregistration[all]

Minimal example

from polar import load_sample_data, POLAR

X, degradations, R_abs_gt = load_sample_data()

model = POLAR(**degradations)
X_hat = model.fit_transform(X)

POLAR partially respects the Scikit-Learn Estimator API. Specifically, it has the two main methods:

• fit(X: Tensor | Sequence[Tensor]) -> None
• fit_transform(X: Tensor | Sequence[Tensor]) -> Tensor | Sequence[Tensor]

They take a list of PyTorch tensors (each of shape \((K_i, 3)\) if point clouds are of varying lengths) or a single batch tensor of shape \((N, K, 3)\) containing all the (same length) views.