search

fabiopoiesi / gedi

Project page of the paper "Learning general and distinctive 3D local deep descriptors for point cloud registration" published in IEEE T-PAMI
131 stars 11 forks source link

readme

GeDi: Learning general and distinctive 3D local deep descriptors for point cloud registration - IEEE T-PAMI

Official repository of GeDi descriptor. Paper (pdf)

PWC PWC PWC PWC PWC

3DMatch ⟶ ETH 3DMatch ⟶ KITTI KITTI ⟶ 3DMatch

An effective 3D descriptor should be invariant to different geometric transformations, such as scale and rotation, robust to occlusions and clutter, and capable of generalising to different application domains. We present a simple yet effective method to learn general and distinctive 3D local descriptors (GeDi) that can be used to register point clouds that are captured in different domains. Point cloud patches are extracted, canonicalised with respect to their local reference frame, and encoded into scale and rotation-invariant compact descriptors by a deep neural network that is invariant to permutations of the input points. This design is what enables GeDi to generalise across domains.

Gedi is an extension of DIP descriptor. Additional code and data can be found in DIP repository.

Tested with

Installation

Set up your environment and start it 

virtualenv venv -p python3
source venv/bin/activate
pip install --upgrade pip

In order to have Open3D functioning correctly on GPU, you should download and install this version of torch (link). Once you have downloaded it, install it as

pip install torch-1.8.1-cp38-cp38-linux_x86_64.whl

Then install the following packages

pip install open3d==0.15.1
pip install torchgeometry==0.1.2
pip install gdown
pip install tensorboard
pip install protobuf==3.20

Note: the last pip is to downgrade protobuf in order to avoid your system complaining that the version is too high.

Now, you have to compile and install Pointnet2. This may be a bit tricky, as it could complain if the version of CUDA is different from that of PyTorch. You can check the original repository for more information.

cd backbones
pip install ./pointnet2_ops_lib/

Download data

The script download_data.py will download the pretained model and assets to run the demo. Data will be put in the right directories automatically. The model was trained on 3DMatch training set.

python download_data.py

Demo

Once the data are downloaded, execute the demo as

python demo.py

The result will look like these (note that results may slightly differ from run to run due to the randomisation of RANSAC).

Before registration After registration

Citing our work

Please cite the following paper if you use our code

@inproceedings{Poiesi2021,
  title = {Learning general and distinctive 3D local deep descriptors for point cloud registration},
  author = {Poiesi, Fabio and Boscaini, Davide},
  booktitle = {IEEE Trans. on Pattern Analysis and Machine Intelligence},
  year = {(early access) 2022}
}

Acknowledgements

This research was supported by the SHIELD project, funded by the European Union’s Joint Programming Initiative – Cultural Heritage, Conservation, Protection and Use joint call (link), and partially by Provincia Autonoma di Trento (Italy) under L.P. 6/99 as part of the X-Loader4.0 project (link).

TODOS