SUPERSLAM3 is a visual odometry pipeline that combines the Superpoint-based front end with the ORB-SLAM3 pose estimation backend. In the implemented pipeline, traditional ORB and BRIEF feature detection and description methods have been replaced with the Superpoint pipeline.
In the SUPERSLAM3 pipeline, input images are converted to grayscale and fed into the Superpoint detector pipeline (A). The Superpoint encoder-decoder pipeline consists of a learned encoder, utilizing several convolutional layers, and two non-learned decoders for joint feature and descriptor extraction. The detected features are then processed by the ORB-SLAM3 backend, which comprises three primary components operating in parallel threads: the Tracking, Local Mapping, and Loop & Map Merging threads (B). The backend extracts keyframes, initializes and updates the map, and performs both local and global motion and pose estimation within the Local Mapping Thread and Loop & Map Merging thread. If a loop closure is detected, the pose estimation is further refined.
This repository was forked from ORB-SLAM3. The pre-trained model of SuperPoint come from the official MagicLeap repository.
7, September 2023 update
- SUPERSLAM3 v1.0 is now publicly available!
- REPOSITORY UNDER CONSTRUCTION: We are in the process of uploading and building the GitHub repository.
- We are currently testing the project on Ubuntu 20.04 and 22.02 with upgraded CUDA, CuDNN, and libtorch libraries.
[ORB-SLAM3] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel and Juan D. Tardós, ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM, IEEE Transactions on Robotics 37(6):1874-1890, Dec. 2021. PDF.
[Superpoint] DeTone, Daniel, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2018, PDF.
[DBoW2 Place Recognition] Dorian Gálvez-López and Juan D. Tardós. Bags of Binary Words for Fast Place Recognition in Image Sequences. IEEE Transactions on Robotics, vol. 28, no. 5, pp. 1188-1197, 2012, PDF.
[SuperPoint-SLAM] Deng, Chengqi, et al. Comparative study of deep learning based features in SLAM. 2019 4th Asia-Pacific Conference on Intelligent Robot Systems (ACIRS). IEEE, 2019, PDF.
If you use SUPERSLAM3 in an academic work, please cite:
@article{mollica2023integrating,
title={Integrating Sparse Learning-Based Feature Detectors into Simultaneous Localization and Mapping—A Benchmark Study},
author={Mollica, Giuseppe and Legittimo, Marco and Dionigi, Alberto and Costante, Gabriele and Valigi, Paolo},
journal={Sensors},
volume={23},
number={4},
pages={2286},
year={2023},
publisher={MDPI}
}
We have tested the libraries and executables on Ubuntu 18.04.
ORBSLAM3 uses the new thread and chrono functionalities of C++11.
We use OpenCV to manipulate images and features. Dowload and install instructions can be found at: http://opencv.org. Required at least 3.0. Tested with OpenCV 3.4.11.
Required by g2o (see below). Download and install instructions can be found at: http://eigen.tuxfamily.org. Required at least 3.1.0. Tested with Eigen3 3.4.0.
We use a BOW vocabulary based on the BOW3 library to perform place recognition, and g2o library is used to perform non-linear optimizations. All these libraries are included in the Thirdparty folder.
The DBOW3 vocabulary can be downloaded from google drive. Place the uncompressed vocabulary file into the Vocabulary
directory within the SUPERSLAM3 project.
For more informations please refer to this repo.
Please, follow these instructions for the installation of the Cuda Toolkit 10.2.
If not installed during the Cuda Toolkit installation process, please install the nvidia driver 440:
sudo apt-get install nvidia-driver-440
Export Cuda paths
echo 'export PATH=/usr/local/cuda-10.2/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc
sudo ldconfig
Verify the Nvidia driver availability:
nvidia-smi
Download CuDNN 7.6.5 from the official NVidia page, and install the headers and libraries in the local CUDA installation folder:
sudo cp -P <PATH_TO_CUDNN_FOLDER>/cuda/include/cudnn.h <PATH_TO_CUDA10.1_FOLDER>/include/
sudo cp -P <PATH_TO_CUDNN_FOLDER>/cuda/lib64/libcudnn* <PATH_TO_CUDA10.1_FOLDER>/lib64/
sudo chmod a+r <PATH_TO_CUDA10.2_FOLDER>/lib64/libcudnn*
The CUDA installation can be verified by running:
nvcc -V
If only CPU can be used, install cpu-version LibTorch. Some code change about tensor device should be required.
wget -O LibTorch.zip https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.6.0.zip
sudo unzip LibTorch.zip -d /usr/local
Clone the repository:
git clone --recursive https://github.com/isarlab-department-engineering/SUPERSLAM3
Use the provided script build.sh
to build the Thirdparty libraries and SUPERSLAM3 project. Please make sure you have installed all required dependencies (see section 1).
Open the build.sh file and modify the weights file path according to the absolute path of the weights file in your PC (<PATH_TO_SUPERSLAM3_FOLDER>/Weights/superpoint.pt)
cmake .. -DCMAKE_BUILD_TYPE=Release -DSUPERPOINT_WEIGHTS_PATH="<PATH_TO_SUPERSLAM3_FOLDER>/Weights/superpoint.pt"
Build the project:
cd <PATH_TO_SUPERSLAM3_FOLDER>
chmod +x build.sh
./build.sh
To test SUPERSLAM3 with the EUROC dataset:
-
Download the MH01 sequence (ASL Dataset format) from this link.
-
Unzip the downloaded sequence and execute the following command. Change PATH_TO_MH01_SEQUENCE_FOLDER to the uncompressed dataset folder.
cd <PATH_TO_SUPERSLAM3_FOLDER>
./Examples/Monocular/mono_euroc ./Vocabulary/superpoint_voc.yml ./Examples/Monocular/EuRoC.yaml <PATH_TO_MH01_SEQUENCE_FOLDER> ./Examples/Monocular/EuRoC_TimeStamps/MH01.txt