DeepUME: Learning the Universal Manifold Embedding for Robust Point Cloud Registration

[video] [paper] [supplementary] [data] [thesis]

Introduction

Deep Universal Manifold Embedding (DeepUME) is a learning-based point cloud registration algorithm which achieves fast and accurate global regitration. This repository contains a basic PyTorch implementation of DeepUME. Please refer to our paper for more details.

Usage

This code has been tested on Python 3.6.13, PyTorch 1.4.0 and CUDA 10.1.

Prerequisite

1. PyTorch=1.4.0: https://pytorch.org
2. h5py
3. open3d
4. TensorboardX: https://github.com/lanpa/tensorboardX
5. Download data to data/.

Training

python main.py --exp_name=deepume --noise=sampling

Testing

python main.py --exp_name=deepume --eval 
or
python main.py --exp_name=pretrained --eval --pretrained='pretrained/deepume.t7' --noise=zero_intersec --test_dataset=FAUST