C++ implementation of SuperPoint inference in LibTorch

This implements SuperPoint: Self-Supervised Interest Point Detection and Description in C++ using LibTorch and TorchScript.

matches	tracks

Installation

This C++ implementation of SuperPoint requires LibTorch. The following installation instructions will use vcstool, rosdep and colcon for managing source and binary dependencies and to create a relocatable workspace.

CUDA and cuDNN:
- add the official CUDA Toolkit repo for Ubuntu
- install CUDA and cuDNN runtime sudo apt install cuda-runtime-11-3 libcudnn8

vcstool, rosdep and colcon:

sudo pip3 install -U vcstool rosdep colcon-common-extensions
sudo rosdep init
rosdep update

create and compile the colcon workspace:

# create workspace folder
mkdir ~/super_point_ws/
cd ~/super_point_ws/
# download sources
vcs import << EOF
- git: {local-name: src/torch_cpp,             uri: "https://github.com/christian-rauch/torch_cpp.git"}
- git: {local-name: src/super_point_inference, uri: "https://github.com/christian-rauch/super_point_inference.git"}
EOF
# resolve binary dependencies
rosdep install --from-paths src --ignore-src -y
# build workspace
colcon build --cmake-args -D CMAKE_BUILD_TYPE=Release -D CMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc

The torch_cpp and super_point_inference packages will be installed to ~/super_point_ws/install/. After sourcing the workspace via source ~/super_point_ws/install/setup.bash those packages will be discoverable by CMake.

Keypoint Extraction and Matching Example

The super_point_inference contains the original model and weights converted via TorchScript and an example program to extract keypoints and find matches in a sequence of images. The converted model and weights are located at ~/super_point_ws/install/super_point_inference/share/weights/SuperPointNet.pt. The example program superpoint_match takes the path to the converted model file and a list of image paths as argument:

# download example data
wget https://raw.githubusercontent.com/magicleap/SuperPointPretrainedNetwork/master/assets/icl_snippet/{250,254,258}.png -P /tmp/images/
# source the workspace
source ~/super_point_ws/install/setup.bash
# extract keypoints and matches
superpoint_match ~/super_point_ws/install/super_point_inference/share/weights/SuperPointNet.pt /tmp/images/{250,254,258}.png

The matches between consecutive image pairs will be shown in separate windows (press any key to close) and exported as /tmp/matches-$PAIR.png and /tmp/tracks-$PAIR.png.

Citation

If you use the original work by Daniel DeTone et al., please cite their work:

@InProceedings{DeTone2018,
  author =    {DeTone, Daniel and Malisiewicz, Tomasz and Rabinovich, Andrew},
  title =     {SuperPoint: Self-Supervised Interest Point Detection and Description},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year =      {2018}
}

If you use this implementation, please cite our work:

@ARTICLE{Rauch2022,
  author={Rauch, Christian and Long, Ran and Ivan, Vladimir and Vijayakumar, Sethu},
  journal={IEEE Robotics and Automation Letters},
  title={Sparse-Dense Motion Modelling and Tracking for Manipulation Without Prior Object Models},
  year={2022},
  volume={7},
  number={4},
  pages={11394-11401},
  doi={10.1109/LRA.2022.3200177}
}