Pytorch Chamfer Distance.

Include a CUDA version, and a PYTHON version with pytorch standard operations. NB : In this depo, dist1 and dist2 are squared pointcloud euclidean distances, so you should adapt thresholds accordingly.

F - Score

CUDA VERSION

JIT compilation
Supports multi-gpu
2D point clouds.
3D point clouds.
5D point clouds.
Contiguous() safe.

Python Version

Supports any dimension

Usage

import torch, chamfer3D.dist_chamfer_3D, fscore
chamLoss = chamfer3D.dist_chamfer_3D.chamfer_3DDist()
points1 = torch.rand(32, 1000, 3).cuda()
points2 = torch.rand(32, 2000, 3, requires_grad=True).cuda()
dist1, dist2, idx1, idx2 = chamLoss(points1, points2)
f_score, precision, recall = fscore.fscore(dist1, dist2)

Add it to your project as a submodule

git submodule add https://github.com/ThibaultGROUEIX/ChamferDistancePytorch

Benchmark: [forward + backward] pass

CUDA 10.1, NVIDIA 435, Pytorch 1.4
p1 : 32 x 2000 x dim
p2 : 32 x 1000 x dim

Timing (sec 1000)*	2D	3D	5D
Cuda Compiled	1.2	1.4	1.8
Cuda JIT	1.3	1.4	1.5
Python	37	37	37

Memory (MB)	2D	3D	5D
Cuda Compiled	529	529	549
Cuda JIT	520	529	549
Python	2495	2495	2495

What is the chamfer distance ?

Stanford course on 3D deep Learning

Aknowledgment

Original backbone from Fei Xia.

JIT cool trick from Christian Diller

Troubleshoot

Undefined symbol: Zxxxxxxxxxxxxxxxxx :

--> Fix: Make sure to import torch before you import chamfer. --> Use pytorch.version >= 1.1.0

RuntimeError: Ninja is required to load C++ extension

wget https://github.com/ninja-build/ninja/releases/download/v1.8.2/ninja-linux.zip
sudo unzip ninja-linux.zip -d /usr/local/bin/
sudo update-alternatives --install /usr/bin/ninja ninja /usr/local/bin/ninja 1 --force

TODO:

Discuss behaviour of torch.min() and tensor.min() which causes issues in some pytorch versions

2.5 KiB Raw Blame History