NOTE: This implementation was stolen from the pytorch3d repo, and all I did was to simply repackage it.
A simple example Pytorch module to compute Chamfer distance between two pointclouds.
You can install the package using pip
.
pip install chamferdist
In your favourite python/conda virtual environment, execute the following commands.
NOTE: This assumes you have PyTorch installed already (preferably, >= 1.5.0; untested for earlier releases).
python setup.py install
That's it! You're now ready to go. Here's a quick guide to using the package. Fire up a terminal. Import the package.
>>> import torch
>>> from chamferdist import ChamferDistance
Create two random pointclouds. Each pointcloud is a 3D tensor with dimensions batchsize
x number of points
x number of dimensions
.
>>> source_cloud = torch.randn(1, 100, 3).cuda()
>>> target_cloud = torch.randn(1, 50, 3).cuda()
Initialize a ChamferDistance
object.
>>> chamferDist = ChamferDistance()
Now, compute Chamfer distance.
>>> dist_forward = chamferDist(source_cloud, target_cloud)
>>> print(dist_forward.detach().cpu().item())
Here, dist
is the Chamfer distance between source_cloud
and target_cloud
. Note that Chamfer distance is not bidirectional (and, in stricter parlance, it is not a distance metric).
The Chamfer distance in the backward direction, i.e., target_cloud
to source_cloud
can be computed in two ways. The naive way is to simply flip the order of the arguments, i.e.,
>>> dist_backward = chamferDist(target_cloud, source_cloud)
Another way is to use the reverse
flag provided by the ChamferDistance
module, i.e.,
>>> dist_backward = chamferDist(source_cloud, target_cloud, reverse=True)
>>> print(dist_backward.detach().cpu().item())
Typically, a symmetric version of the Chamfer distance is obtained, by summing the "forward" and the "backward" Chamfer distances. This is supported by the bidirectional
flag.
>>> dist_bidirectional = chamferDist(source_cloud, target_cloud, bidirectional=True)
>>> print(dist_bidirectional.detach().cpu().item())
Look at the example script for more details: example.py
If you find this work useful, you might want to cite the original implementation from which this codebase was borrowed (stolen!) - PyTorch3D.
@article{ravi2020pytorch3d,
author = {Nikhila Ravi and Jeremy Reizenstein and David Novotny and Taylor Gordon
and Wan-Yen Lo and Justin Johnson and Georgia Gkioxari},
title = {Accelerating 3D Deep Learning with PyTorch3D},
journal = {arXiv:2007.08501},
year = {2020},
}