Code has been added for the most recent paper based on SLAMBench:
Robust SLAM Systems: Are We There Yet?
https://robustslam.github.io/evaluation
- What is SLAMBench?
- How to set up SLAMBench?
- What algorithms does SLAMBench support?
- How to run an existing algorithm with SLAMBench?
- How to add a new benchmark in SLAMBench?
- Known Issues
- Release History
Use the following command to list all available algorithms:
make usecases
SLAMBench is a SLAM performance benchmark that combines a framework for quantifying quality-of-result with instrumentation of accuracy, execution time, memory usage and energy consumption. It also include a graphical interface to visualize these information.
SLAMBench offers a platform for a broad spectrum of future research in jointly exploring the design space of algorithmic and implementation-level optimisations. It targets desktop, laptop, mobile and embedded platforms. Some of the benchmarks (in particular KFusion) were tested on Ubuntu, OS X and Android (more information about android here https://github.com/bbodin/slambench-android).
SLAMBench currently supports the following algorithms:
- ORB-SLAM3 [Campos et al, ARXIV'20]: C++ as distributed by https://github.com/UZ-SLAMLab
- ReFusion [Palazollo et al. IROS'19]: CUDA as distributed by https://github.com/PRBonn
- OpenVINS [Geneva et al. IROS'19]: C++ as distributed by https://github.com/rpng/
- Supereight [Vespa et al. RA-L'18]: C++, OpenMP as distributed by https://github.com/emanuelev
- BundleFusion [Dai et al. ACM TOG'17]: CUDA as distributed by https://github.com/niessner
- SemanticFusion [McCormac et al. ICRA'17]: CUDA as distributed by https://github.com/seaun163
- ORB-SLAM2 [Mur-Artal et al, TOR'15 and TOR'17]: C++ as distributed by https://github.com/raulmur
- DSO [Engel et al. Arxiv'16]: C++ as distributed by https://github.com/JakobEngel
- ElasticFusion [Whelan et al, IJRR'16]: CUDA as distributed by https://github.com/mp3guy
- InfiniTAMv2 [Kahler et al, ISMAR'15]: C++, OpenMP and CUDA versions as distributed by https://github.com/victorprad/
- KinectFusion [Newcombe et al. ISMAR'11]: C++, OpenMP, OpenCL and CUDA inspired by https://github.com/GerhardR
- LSDSLAM [Engel et al, ECCV'14]: C++, and threaded as distributed by https://github.com/tum-vision/ and modified by https://github.com/mp3guy
- MonoSLAM [Davison et al, TPAMI'07]: Original version as distributed by https://github.com/hanmekim/
- OKVIS [Leutenegger et al, IJRR'15]: Original version as distributed by https://github.com/ethz-asl
- PTAM [Klein et al, ISMAR'07 and ECCV'08]: Original version as distributed by https://github.com/Oxford-PTAM/
- SVO [Forster et al, ICRA'14]: Original version as distributed by https://github.com/uzh-rpg/rpg_svo/ (a more recent version available at http://rpg.ifi.uzh.ch/svo2.html)
IMPORTANT: If you use any of those algorithms in scientific publications, you should refer to the respective publications.
In addition, if you use SLAMBench in scientific publications, we would appreciate citations to the following papers:
@inproceedings{bujanca2021robust,
author={Bujanca, Mihai and Shi, Xuesong and Spear, Matthew and Zhao, Pengpeng and Lennox, Barry and Luj{\'a}n, Mikel},
booktitle={2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
title={Robust SLAM Systems: Are We There Yet?},
year={2021},
doi={10.1109/IROS51168.2021.9636814}
}
@inproceedings{bujanca2019slambench,
title={SLAMBench 3.0: Systematic automated reproducible evaluation of SLAM systems for robot vision challenges and scene understanding},
author={Bujanca, Mihai and Gafton, Paul and Saeedi, Sajad and Nisbet, Andy and Bodin, Bruno and O'Boyle, Michael FP and Davison, {Andrew J} and Kelly, {Paul H.J.} and Riley, Graham and Lennox, Barry and Luj{\'a}n, Mikel and Furber, Steven},
booktitle={2019 International Conference on Robotics and Automation (ICRA)},
pages={6351--6358},
year={2019},
organization={IEEE}
}
@inproceedings{Bodin2018,
author = "Bruno Bodin and Harry Wagstaff and Sajad Saeedi and Luigi Nardi and Emanuele Vespa and Mayer, {John H} and Andy Nisbet and Mikel Luj{\'a}n and Steve Furber and Davison, {Andrew J} and Kelly, {Paul H.J.} and Michael O'Boyle",
title = "SLAMBench2: Multi-Objective Head-to-Head Benchmarking for Visual SLAM",
booktitle = "{IEEE Intl. Conf. on Robotics and Automation (ICRA)}",
year = {2018},
month = {May}
}
@inproceedings{Nardi2015,
title={Introducing SLAMBench, a performance and accuracy benchmarking methodology for SLAM},
author={Nardi, Luigi and Bodin, Bruno and Zia, M Zeeshan and Mawer, John and Nisbet, Andy and Kelly, Paul HJ and Davison, Andrew J and Luj{\'a}n, Mikel and O'Boyle, Michael FP and Riley, Graham and others},
booktitle={2015 IEEE international conference on robotics and automation (ICRA)},
pages={5783--5790},
year={2015},
organization={IEEE}
}
As SLAMBench deals with multiple SLAM algorithms, dependencies might be difficult to install on any systems. To ease the usage of SLAMBench we provide auto-installation of dependencies and recommend the use fresh installation of Ubuntu 18/20 or Fedora 24-29.
- CMake 2.8.11 or higher is required.
- Make
- GCC C/C++
- Boost (Optional)
- GLUT (Optional)
- Git
- Mercurial
- wget
- unzip
- lapack
- blas
- findutils
- cvs
- glog
- gflags
- p7zip
With Fedora 29: dnf install -y yaml-cpp-devel gtk2-devel mesa-libEGL-devel vtk-devel cmake make git mercurial wget unzip gcc gcc-c++ lapack blas lapack-devel blas-devel findutils cvs glut-devel glew-devel boost-devel glog-devel gflags-devel libXmu-devel p7zip
With Fedora 24:
dnf install -y gtk2-devel vtk-devel cmake make git mercurial wget unzip gcc gcc-c++ lapack blas lapack-devel blas-devel findutils cvs glut-devel glew-devel boost-devel glog-devel gflags-devel libXmu-devel
With Ubuntu 20.04:
apt-get -y install libvtk6.3 libvtk6-dev unzip libflann-dev wget mercurial git gcc g++ cmake python-numpy freeglut3 freeglut3-dev libglew-dev libglu1-mesa libglu1-mesa-dev libgl1-mesa-glx libgl1-mesa-dev libxmu-dev libxi-dev libboost-all-dev cvs libgoogle-glog-dev libatlas-base-dev gfortran gtk2.0 libgtk2.0-dev libyaml-dev build-essential libyaml-cpp-dev
With Ubuntu 18.04:
apt-get -y install libvtk6.3 libvtk6-dev unzip libflann-dev wget mercurial git gcc g++ cmake python-numpy freeglut3 freeglut3-dev libglew-dev libglu1-mesa libglu1-mesa-dev libgl1-mesa-glx libgl1-mesa-dev libxmu-dev libxi-dev libboost-all-dev cvs libgoogle-glog-dev libatlas-base-dev gfortran gtk2.0 libgtk2.0-dev libyaml-dev build-essential libyaml-cpp-dev
To run the CUDA implementation of some of the algorithms, you will need extra dependencies.
With Ubuntu: apt-get -y install nvidia-cuda-toolkit clinfo
With Fedora: yum install cuda
Install dependencies first [NOTE can be installed by the user on its system as well]:
make deps
The idea is to maximise the chance of a good build, by selection the best cocktail of libraries. This will download and compile the following applications: Brisk, Ceres, CVD, Eigen3, Flann, FreeImage, G2O, Gvars, OpenCV, OpenGV, OpenTuner, Pangolin, PCL, Suitesparse, TooN.
You can also install each ofthese individually, using the commands such as: eigen3
, flann
, g2o
, opencv
, opengv
, pcl
, toon
, suitesparse
, ...
more information is available in the framework/makefiles/README.md
file.
SLAMBench is a framework that can be compiled by simply running:
make slambench
Although, by doing this, you only compile the libraries of SLAMBench.
To download use-cases, there are specific target named after the type of algorithm you need to test:
make kfusion lsdslam
Then to compile these specific use-case, you will need to specify identifiers together with the slambench target:
make slambench APPS=kfusion,lsdslam
The current benchmarks identifiers are efusion, infinitam, kfusion, lsdslam, monoslam, okvis, ptam, orbslam2, svo.
You will find more information to download and compile use-cases with the make usecases
command.
To test a SLAM algorithm you can use a Live camera, or a dataset. SLAMBench provides tools to automatically download some of the most popular datasets, that is ICL-NUIM and TUM RGB-D. The file format (*.slam) will then include all the most important information about the dataset, those are Camera calibration setting, initial position of the sensors, and the ground truth.
As an example to download and generate the Living Room Trajectory 2 from the ICLNUIM dataset, you can run the following :
> make datasets/ICL_NUIM/living_room_traj2_loop.slam
SLAMBench currently supports the following datasets:
- OpenLORIS [Shi et al, ICRA'20]: Lifelong SLAM dataset
- Bonn Dynamic [Palazollo et al. IROS'19]: Dynamic scene dataset
- UZH-FPV [Delmerico et al. ICRA'19]: Drone racing dataset
- ETH Illumination [Park et al, ICRA'17]: Illumination changes dataset
- VolumeDeform [Innmann et al, ECCV'16]: Non-rigid reconstruction
- EuRoC MAV [Burri et al, IJRR'16]: Micro Aerial Vehicle dataset
- ICL-NUIM [Handa et al, ICRA'14]: Synthetic dataset
- TUM RGB-D [Sturm et al, IROS'12]: A standard SLAM benchmark
A complete list of the datasets available is provided by the command
make datasets
.
SLAMBench is already compatible with a wide range of algorithms which are not included in this repository (see above for list of algorithms).
However you can easily integrate those algorithms using the command:
make usecases
This command will explain in details how to integrate algorithms that are already compatible with SLAMBench.
Once you have compiled a benchmark, there are several ways to run it.
For each implementation of this benchmark, you will find a specific library.
As an example, with KinectFusion, after running make slambench APPS=kfusion
, you may find the following libraries in the build/lib
directory :
> ls build/lib/libkfusion-*-library.so
build/lib/libkfusion-cpp-library.so
build/lib/libkfusion-notoon-library.so
build/lib/libkfusion-openmp-library.so
build/lib/libkfusion-cuda-library.so
build/lib/libkfusion-opencl-library.so
We can see five different implementations (cpp, notoon, and openmp, cuda and opencl). The list of available binaries depends of the dependencies you installed beforehand. For example, you need CUDA to compile the kfusion-cuda. A complete list of the dependencies is available at the end of this README.
To run one algorithm you will need to use a loader. There are three different loaders supported, benchmark, pangolin, and lifelong. The first two loaders are used the same way, except that benchmark is a command line application dedicated to measurements, while pangolin is a graphical user interface less precise in term of measurement but which provide a good interface for demonstrations. The lifelong loader can take multiple input (multiple .slam files following the -i option, separated by ',') which will be sent to the benchmark one by one. Other than that it is similar to the benchmark loader. There is currently no loader both supporting loading multiple input and having a graphical user interface.
Each loader has a series of parameters to specify such as the dataset location, or the libraries to run. The list of those parameters is available by using the "--help" parameters.
> ./build/bin/benchmark_loader --help
== SLAMBench Configuration ==
Available parameters :
-fl --frame-limit : last frame to compute (Default=0)
-o --log-file : Output log file (Default=)
-i --input : Specify the input file or mode. (Default=)
-load --load-library : Load a specific SLAM library. (Default=)
-dse --dse : Output solution space of parameters. (Default=false)
-h --help : Print the help. (Default=false)
-nf --negative-focal-length : negative focal length (Default=false)
-realtime --realtime-mode : realtime frame loading mode (Default=false)
-realtime-mult --realtime-multiplier : realtime frame loading mode (Default=1)
-fo --file-output : File to write slamfile containing outputs (Default=)
Then if you run the loader again, while providing a dataset file -i dataset.slam
, you will see new parameters dedicated to the dataset:
> ./build/bin/benchmark_loader -i datasets/ICL_NUIM/living_room_traj2_loop.slam --help
== SLAMBench Configuration ==
Available parameters :
....
-Camera-intrisics --Camera-intrisics : (Default=nullptr Current=0.751875,1,0.4992185,0.4989583)
-Depth-intrisics --Depth-intrisics : (Default=nullptr Current=0.751875,1,0.4992185,0.4989583)
-Depth-dip --Depth-disparity-params : (Default=nullptr Current=0.001,0)
-Camera-intrisics --Camera-intrisics : (Default=nullptr Current=0.751875,1,0.4992185,0.4989583)
Finally is you add a library name -load libname
, more parameter can be seen:
> ./build/bin/benchmark_loader -i datasets/ICL_NUIM/living_room_traj2_loop.slam -load ./build/lib/libkfusion-cpp-library.so --help
== SLAMBench Configuration ==
Available parameters :
....
-c --compute-size-ratio : Compute ratio (Default=1)
-r --integration-rate : integration-rate (Default=2)
-t --tracking-rate : tracking-rate (Default=1)
-z --rendering-rate : rendering-rate (Default=4)
-l --icp-threshold : icp-threshold (Default=1e-05)
-m --mu : mu (Default=0.1)
-s --volume-size : volume-size (Default=8,8,8)
-d --volume-direction : volume-direction (Default=4,4,4)
-v --volume-resolution : volume-resolution (Default=256,256,256)
-y1 --pyramid-level1 : pyramid-level1 (Default=10)
-y2 --pyramid-level2 : pyramid-level2 (Default=5)
-y3 --pyramid-level3 : pyramid-level3 (Default=4)
You can run a loader with only one dataset at a time and it must be specified first.
In the next section we will explain how to use SLAMBench to evaluate the performance of a SLAM algorithm.
SLAMBench works with Metrics and Outputs elements.
When you run the benchmark_loader
or the pangolin_loader
or the lifelong_loader
these are those elements that you can visualize.
Metrics are components generated by SLAMBench framework really, while Outputs are generated by the algorithm or may be elements post-processed by SLAMBench (such as the aligned trajectory with the ground truth).
Let us run the benchmark loader. Its output is composed of two main parts, the Properties
section, and the Statistics
section.
the properties section details all the parameters used for the experiment (could been changed or not via the command line).
the statistics section report all the outputs and metrics selection for output in the benchmark loader.
> ./build/bin/benchmark_loader -i datasets/ICL_NUIM/living_room_traj2_loop.slam -load ./build/lib/libkfusion-cpp-library.so
SLAMBench Report run started: 2018-02-02 04:41:31
Properties:
=================
frame-limit: 0
log-file:
input: datasets/ICL_NUIM/living_room_traj2_loop.slam
load-library: ./build/lib/libkfusion-cpp-library.so
dse: false
help: false
negative-focal-length: false
realtime-mode: false
realtime-multiplier: 1
file-output:
Camera-intrisics: 0.751875,1,0.4992185,0.4989583
Depth-intrisics: 0.751875,1,0.4992185,0.4989583
Depth-disparity-params: 0.001,0
Camera-intrisics: 0.751875,1,0.4992185,0.4989583
compute-size-ratio: 1
integration-rate: 2
tracking-rate: 1
rendering-rate: 4
icp-threshold: 1e-05
mu: 0.1
volume-size: 8,8,8
volume-direction: 4,4,4
volume-resolution: 256,256,256
pyramid-level1: 10
pyramid-level2: 5
pyramid-level3: 4
Statistics:
=================
Frame Number Timestamp Duration_Frame GPU_Memory CPU_Memory Duration_Preprocessing Duration_Tracking Duration_Integration Duration_Raycasting Duration_Render X Y ZATE_Frame
1 0.0000000000 0.7679200000 0 623801799 0.1254800000 0.0195420000 0.0561620000 0.0000030000 0.5667170000 4.0000000000 4.0000000000 4.0000000000 0.0000002980
2 1.0000000000 0.2003970000 0 623801799 0.1242030000 0.0156470000 0.0581670000 0.0000000000 0.0023710000 4.0000000000 4.0000000000 4.0000000000 0.0010031639
3 2.0000000000 0.1989980000 0 623801799 0.1233680000 0.0152360000 0.0580180000 0.0000000000 0.0023690000 4.0000000000 4.0000000000 4.0000000000 0.0055015362
4 3.0000000000 0.7518580000 0 623801799 0.1220660000 0.0152080000 0.0563070000 0.5559520000 0.0023170000 4.0000000000 4.0000000000 4.0000000000 0.0036504765
5 4.0000000000 1.3683420000 0 623801799 0.1240890000 0.0767240000 0.0581630000 0.5504240000 0.5589330000 3.9957129955 4.0020360947 4.0009112358 0.0021276891
...
The main reason to provide a new version of SLAMBench is not only because of the introduction of new benchmarks but also because we provide now a clear and specific API for SLAM algorithms to be implemented in order to add a new algorithm.
bool sb_new_slam_configuration(SLAMBenchLibraryHelper * slam_settings);
bool sb_init_slam_system(SLAMBenchLibraryHelper * slam_settings);
bool sb_update_frame(SLAMBenchLibraryHelper * slam_settings, slambench::io::SLAMFrame * type);
bool sb_process_once(SLAMBenchLibraryHelper * slam_settings);
bool sb_relocalize(SLAMBenchLibraryHelper * slam_settings);
bool sb_update_outputs(SLAMBenchLibraryHelper *lib, const slambench::TimeStamp *latest_output);
bool sb_clean_slam_system();
bool sb_update_outputs(SLAMBenchUI *);
If each of those functions are correctly implemented for a specific implementation of a specific algorithm, then this algorithm is compatible with SLAMBench and can be evaluated as well.
In this section we will present those functions one by one.
This function is called first, and only once, SLAM systems is expected to provide its parameters.
Example :
bool sb_new_slam_configuration(SLAMBenchLibraryHelper * slam_settings) {
slam_settings->addParameter(TypedParameter<float>("c", "confidence", "Confidence", &confidence, &default_confidence));
slam_settings->addParameter(TypedParameter<float>("d", "depth", "Depth", &depth, &default_depth));
slam_settings->addParameter(TypedParameter<int> ("td", "textureDim", "textureDim", &textureDim, &default_textureDim));
return true;
}
should always return true
or an exception will be raised.
This function is called second, and only once, SLAM systems is expected to allocate memory, retrieve sensor informations.
To retrieve sensor there is SensorFinder
:
slambench::io::CameraSensorFinder sensor_finder;
auto rgb_sensor = sensor_finder.FindOne(slam_settings->get_sensors(), {{"camera_type", "rgb"}});
SLAM systems are also expected to define there output, there is one mandatory output, the pose:
pose_output = new slambench::outputs::Output("Pose", slambench::values::VT_POSE, true);
slam_settings->GetOutputManager().RegisterOutput(pose_output);
should always return true
or an exception will be raised.
Algorithms receive frames ordered by timestamp.
When sb_update_frame
returns false
, sb_update_frame
will be directly called again with the next frame, if it returns true
, sb_process_once
will be called once.
Should always return true
or an exception will be raised.
This is newly introduced to support lifelong SLAM evaluation. It will be called when the input sequence has been switched to the next one. The implementation is expected to explicitly trigger tracking lost and invoke the algorithm's re-localization procedure (if there be). It should return whether the relocalization is sucessful from the algorithm's perspective.
For backward compatibility, this function is allowed to be unimplemented in a benchmark. In such cases, the sb_process_once
function will be called in a re-localization situation.
This function is called last, and only once, SLAM systems is expected to clean everything (free memory).
bool sb_clean_slam_system() {
delete eFusion;
delete inputRGB;
delete inputDepth;
return true;
}
should always return true
or an exception will be raised.
bool sb_update_outputs(SLAMBenchLibraryHelper *slam_settings, const slambench::TimeStamp *timestamp)
The algorithm will return visible outputs (Pose, Point cloud, Frames) as defined by the sb_init_slam_system
function.
Example :
bool sb_update_outputs(SLAMBenchLibraryHelper *lib, const slambench::TimeStamp *ts_p) {
slambench::TimeStamp ts = *ts_p;
if(pose_output->IsActive()) {
// Get the current pose as an eigen matrix
Eigen::Matrix4f mat = eFusion->getCurrPose();
std::lock_guard<FastLock> lock (lib->GetOutputManager().GetLock());
pose_output->AddPoint(ts, new slambench::values::PoseValue(mat));
}
should always return true
or an exception will be raised.
KFusion CUDA requires GCC 4.9 to work. To specify a new gcc compiler for CUDA only, you can use the CUDA_HOST_COMPILER
flag as follows :
make slambench APPS=kfusion CUDA_HOST_COMPILER=$(which gcc-4.9)
Modern O.S. are now using more recent version of this compiler, this may introduce several compatibility issues.
To fix one of them, in the compilation process, when compiling CUDA application we use the -D_GLIBCXX_USE_CXX11_ABI=0
flag.
Version 4.0 (Oct 2021)
- Robustness evaluation
Version 3.0 (May 2019)
- Depth estimation
- Dynamic reconstruction
- Semantic reconstruction
Version 2.0 (Feb 2018)
- This release is a complete new version
Release candidate 1.1 (17 Mar 2015)
- Bugfix : Move bilateralFilterKernel from preprocessing to tracking
- Bugfix : Wrong interpretation of ICP Threshold parameter.
- Esthetic : Uniformisation of HalfSampleRobustImage kernel
- Performance : Change float3 to float4 for the rendering kernels (No effect on OpenCL, but high performance improvement with CUDA)
- Performance : Add a dedicated buffer for the OpenCL rendering
- Feature : Add OSX support
Release candidate 1.0 (12 Nov 2014)
- First public release
Copyright (c) 2014-2021 University of Edinburgh, Imperial College, University of Manchester. Developed in the PAMELA project, EPSRC Programme Grant EP/K008730/1 and
The RAIN Hub, funded by the Industrial Strategy Challenge Fund, part of the UK government’s modern Industrial Strategy. The fund is delivered by UK Research and Innovation and managed by EPSRC [EP/R026084/1].