Repository files navigation

NOTE: This implementation was stolen from thepytorch3d 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 usingpip.

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).

pythonsetup.pyinstall

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.

>>>importtorch>>>fromchamferdistimportChamferDistance

Create two random pointclouds. Each pointcloud is a3D tensor with dimensionsbatchsize xnumber of points xnumber of dimensions.

>>>source_cloud=torch.randn(1,100,3).cuda()>>>target_cloud=torch.randn(1,50,3).cuda()

Initialize aChamferDistance 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 betweensource_cloud andtarget_cloud. Note that Chamfer distance is not bidirectional (and, in stricter parlance, it is not adistance metric).

The Chamfer distance in the backward direction, i.e.,target_cloud tosource_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 thereverse flag provided by theChamferDistance 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 thebidirectional 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 theoriginal 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},}