PCAM: Product of Cross-Attention Matrices for Rigid Registration of Point Clouds
Anh-Quan Cao^1,2, Gilles Puy¹, Alexandre Boulch¹, Renaud Marlet^1,3
1valeo.ai, France and ²Inria, France and ³ENPC, France

If you find this code or work useful, please cite our paper:

@inproceedings{cao21pcam,
  title={{PCAM}: {P}roduct of {C}ross-{A}ttention {M}atrices for {R}igid {R}egistration of {P}oint {C}louds},
  author={Cao, Anh-Quan and Puy, Gilles and Boulch, Alexandre and Marlet, Renaud},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2021},
}

1. This code was implemented with python 3.7, pytorch 1.6.0 and CUDA 10.2. Please install PyTorch.

pip install torch==1.6.0 torchvision==0.7.0

2. A part of the code (voxelisation) is using MinkowskiEngine 0.4.3. Please install it on your system.

sudo apt-get update
sudo apt install libgl1-mesa-glx
sudo apt install libopenblas-dev g++-7
export CXX=g++-7 
pip install -U MinkowskiEngine==0.4.3 --install-option="--blas=openblas" -v

4. Clone this repository and install the additional dependencies:

$ git clone https://github.com/valeoai/PCAM.git
$ cd PCAM/
$ pip install -r requirements.txt

5. Install lightconvpoint [5], which is an early version of FKAConv:

$ pip install -e ./lcp

6. Finally, install pcam:

$ pip install -e ./

You can edit pcam's code on the fly and import function and classes of pcam in other project as well.

Follow the instruction on DGR github repository to download both datasets.

Place 3DMatch in the folder /path/to/pcam/data/3dmatch/, which should have the structure described here.

Place KITTI in the folder /path/to/pcam/data/kitti/, which should have the structure described here.

You can create soft links with the command ln -s if the datasets are stored somewhere else on your system.

For these datasets, we use the same dataloaders as in DGR [1-3], up to few modifications for code compatibility.

Download the dataset here and unzip it in the folder /path/to/pcam/data/modelnet/, which should have the structure described here.

Again, you can create soft links with the command ln -s if the datasets are stored somewhere else on your system.

For this dataset, we use the same dataloader as in PRNet [4], up to few modifications for code compatibility.

Download PCAM pretrained models here and unzip the file in the folder /path/to/pcam/trained_models/, which should have the structure described here.

As we randomly subsample the point clouds in PCAM, there are some slight variations from one run to another. In our paper, we ran 3 independent evaluations on the complete test set and averaged the scores.

We provide two different pre-trained models for 3DMatch: one for PCAM-sparse and one for PCAM-soft, both trained using 4096 input points.

To test the PCAM-soft model, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/3dmatch/soft.yaml

To test the PCAM-sparse model on the test set of , type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/3dmatch/sparse.yaml

As in DGR [1], the results can be improved using different levels of post-processing.

1. Keeping only the pairs of points with highest confidence score (the threshold was optimised on the validation set of 3DMatch).

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/3dmatch/soft_filter.yaml
$ python eval.py with ../configs/3dmatch/sparse_filter.yaml

2. Using in addition the refinement by optimisation proposed by DGR [1].

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/3dmatch/soft_refinement.yaml
$ python eval.py with ../configs/3dmatch/sparse_refinement.yaml

3. Using as well the safeguard proposed by DGR [1].

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/3dmatch/soft_safeguard.yaml
$ python eval.py with ../configs/3dmatch/sparse_safeguard.yaml

Note: For a fair comparison, we fixed the safeguard condition so that it is applied on the same proportion of scans as in DGR [1].

We provide two different pre-trained models for KITTI: one for PCAM-sparse and one for PCAM-soft, both trained using 2048 input points.

To test the PCAM-soft model, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/kitti/soft.yaml

To test the PCAM-sparse model, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/kitti/sparse.yaml

As in DGR [1], the results can be improved by refining the results using ICP.

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/kitti/soft_icp.yaml
$ python eval.py with ../configs/kitti/sparse_icp.yaml 

There exist 3 different variants of this dataset. Please refer to [4] for the construction of these variants.

To test the PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/modelnet/soft.yaml
$ python eval.py with ../configs/modelnet/sparse.yaml

To test the PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/modelnet/soft_unseen.yaml
$ python eval.py with ../configs/modelnet/sparse_unseen.yaml

To test the PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python eval.py with ../configs/modelnet/soft_noise.yaml
$ python eval.py with ../configs/modelnet/sparse_noise.yaml

The models are saved in the folder /path/to/pcam/trained_models/new_training/{DATASET}/{CONFIG}, where {DATASET} is the name of the dataset and {CONFIG} give a description of the PCAM architecture and the losses used for training.

To train a PCAM-soft model, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/3dmatch/soft.yaml

You can then test this new model by typing:

$ python eval.py with ../configs/3dmatch/soft.yaml PREFIX='new_training'

To train a PCAM-sparse model, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/3dmatch/sparse.yaml

Training took about 12 days on a Nvidia Tesla V100S-32GB.

You can then test this new model by typing:

$ python eval.py with ../configs/3dmatch/sparse.yaml PREFIX='new_training'

To train PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/kitti/soft.yaml
$ python train.py with ../configs/kitti/sparse.yaml

Training took about 1 day on a Nvidia GeForce RTX 2080 Ti.

You can then test these new models by typing:

$ python eval.py with ../configs/kitti/soft.yaml PREFIX='new_training'
$ python eval.py with ../configs/kitti/sparse.yaml PREFIX='new_training'

Training PCAM on ModelNet took about 10 hours on Nvidia GeForce RTX 2080.

To train PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/modelnet/soft.yaml NB_EPOCHS=10
$ python train.py with ../configs/modelnet/sparse.yaml NB_EPOCHS=10

You can then test these new models by typing:

$ python eval.py with ../configs/modelnet/soft.yaml PREFIX='new_training'
$ python eval.py with ../configs/modelnet/sparse.yaml PREFIX='new_training'

To train PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/modelnet/soft_unseen.yaml NB_EPOCHS=10
$ python train.py with ../configs/modelnet/sparse_unseen.yaml NB_EPOCHS=10

You can then test these new models by typing:

$ python eval.py with ../configs/modelnet/soft_unseen.yaml PREFIX='new_training'
$ python eval.py with ../configs/modelnet/sparse_unseen.yaml PREFIX='new_training'

To train PCAM models, type:

$ cd /path/to/pcam/scripts/
$ python train.py with ../configs/modelnet/soft_noise.yaml NB_EPOCHS=10
$ python train.py with ../configs/modelnet/sparse_noise.yaml NB_EPOCHS=10

You can then test these new models by typing:

$ python eval.py with ../configs/modelnet/soft_noise.yaml PREFIX='new_training'
$ python eval.py with ../configs/modelnet/sparse_noise.yaml PREFIX='new_training'

[1] Christopher Choy, Wei Dong, Vladlen Koltun. Deep Global Registration, CVPR, 2020.

[2] Christopher Choy, Jaesik Park, Vladlen Koltun. Fully Convolutional Geometric Features. ICCV, 2019.

[3] Christopher Choy, JunYoung Gwak, Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR, 2019.

[4] Yue Wang and Justin M. Solomon. PRNet: Self-Supervised Learning for Partial-to-Partial Registration. NeurIPS, 2019.

[5] Alexandre Boulch, Gilles Puy, Renaud Marlet. FKAConv: Feature-Kernel Alignment for Point Cloud Convolution. ACCV, 2020.

PCAM is released under the Apache 2.0 license.