Posting the way you are using small_gicp

[koide3]

TL; DR: Sharing that you are using small_gicp (and related packages) helps the author receive recognition within his organization and enables continued work on this project.

Long description: The author of this package is NOT a full-time OSS developer but just a researcher who dedicates the majority of his working hours to other assignments. His time for OSS projects is limited, and to sustain his involvement, internal acknowledgment within his organization is crucial. Metrics like the number of stars and forks are often overlooked by executives, who may not grasp their significance. Conversely, stating "this package is used by A/B universities and C/D corporations" can effectively communicate its value. So, if you find this package useful for your work, we kindly ask you to leave a comment in this thread. Just saying hello to the author is appreciated. If you could share about your organization and how you use the package, that would immensely support the author.

[AndreasNagel]

I have not used small_gicp directly, but I used it's predecessor fast_gicp through the hdl_graph_slam package for testing the feasibility of localizing the robot using a tactile sensor array during my master thesis in TalTech university, Tallinn and I believe that same package was used in the robominers project later on by the researchers at TalTech, although that needs confirming. small_gicp is a rather new project, so have not directly used that yet in everyday work. Thank you for your work and sharing it with the world!

[koide3]

@AndreasNagel Thanks for sharing your work! It becomes a big help for me!

[SakodaShintaro]

Hello. I am one of the developers at TIER IV. At TIER IV, we are developing an open source autonomous driving software, Autoware, and we are using ndt_omp in it. It works very stable, and we are grateful that it is easy to make our own changes (such as dynamic map loading and covariance estimation with scores).

Recently, we have been discussing the transition to fast_gicp, and although it is still under consideration, we will also consider this small_gicp. As a first step, (it is a personal work), I have evaluated small_gicp at AWSIM Nishi-Shinjuku and obtained good results. 👍 https://tokumini.hatenablog.com/entry/2024/04/07/110000 (Sorry in Japanese)

We are greatly helped by your code. Thank you for publishing such great software!

[koide3]

@SakodaShintaro Thank a lot for your comment! It is a pleasure to see my packages have been used in real situations.

Regarding the evaluation, I noticed that there was a bug that made termination criteria unchanged with the PCL interface. PR #15 would fix this issue and make comparison fair.

[grischi]

Hello @koide3, I sincerely appreciate your contributions. I have utilized ndt_omp and fast_gicp, and I am currently delving into hdl_global_localization, direct_visual_lidar_calibration, and small_gicp. Your open-source software and scientific publications significantly enrich our research field. Thank you!

[koide3]

@grischi Thanks!!

[hridaybavle]

Hello @koide3,

Amazing work again with the small_gicp. We have been using the hdl_graph_slam as a backbone for our lidar_s_graphs since a while now and the scan matching algorithm uses fast_gicp. I think adding small_gicp to the scan matching will further improve the performance. Thanks again for the amazing work!!

[koide3]

Hi Hriday, Thanks a lot for your kind comment! I hope we can meet again at some conference :)

[KOKIAOKI]

Hello @koide3!
Thank you for providing sophisticated open source code!
Currently I fully replaced pcl::PointCloud with small_gicp::PoincCloud in my point cloud registration code because it allows us to manage points and its covariance at the same time and be parallelized everywhere. Furthermore, thanks to this clear c++ implementation, I understood the simple optimization way in point cloud registration. Regardless of whether this source code is maintained, I will refer to this source code continuously in order to be readable code and enhance point cloud registration.

small_gicp must be the amazing gift for the robotics researchers and developers. I will strongly recommend small_gicp to my coworker and friends.
Thank you again!

[koide3]

Thanks @KOKIAOKI !

[RintoDeVries]

Hi! We are a drone inspection company using small_gicp (and previously fast_gicp) both for lidar odometry as well as icp localization. Thanks for the amazing work!

[koide3]

Thank you @RintoDeVries !

[bpoebiapl]

Hello. I'm a researcher in whereable.ai, an autonomous driving startup in Korea. Your work is amazing and we expect that applying your results to the SLAM framework of our system will lead to significant performance improvements. The relevant industry will make great progress through your results. We would like to thank not only you but also AIST for their support in achieving these results.

[koide3]

Thank you for your comment @bpoebiapl !

[kennyjchen]

Hi @koide3 -- your work has been vital for my PhD research and beyond, including DLO and DLIO. Your hard work and commitment to OSS is deeply appreciated, and my sincerest thanks to you for your contributions to this community!

[koide3]

Thanks @kennyjchen ! I've been following your works, and I'm so glad to see that my package could help you great jobs!

[narutojxl]

Hi Dr. @koide3, I am a robot practitioner. Since of year from 2019 to now, i have learned a lot knowledge from your series excellent works, including hdl_localization, fast_gicp, interactive_slam etc, discussing with you about these repos issues on github. I'm very appreciated, and want to give my sincerest thanks to you. Also thanks Dr. @kennyjchen, i have also learned a lot knowledge from your DLO and DLIO work. I'm much appreciated.

[aserbremen]

Dear @koide3, thanks for this new contribution to the robotics community. Today, I performed a proof of concept for integrating small_gicp into Multi-Robot-Graph-SLAM mrg_slam, the multi-robot SLAM based on hdl_graph_slam.

I ran scan_matching_odometry with fast_gicp and small_gicp using 8 threads and ~4,700 points in the point cloud. fast_gicp took ~5.6 ms per scan and small_gicp ~3.2 ms, which also aligns with with the expected performance increase, which is great. The benefit in the loop detector was not as big but still noticable.

This post might be a little to technical. But could you take a look on how I register small_gicp in mrg_slam here and let me know if I should set more registrations options. The following are the default values:

reg_num_threads: 8
reg_transformation_epsilon: 0.1
reg_maximum_iterations: 64
reg_max_correspondence_distance: 2.0
reg_correspondence_randomness: 20

Edit: We use mrg_slam in multi-robot exploration for ground robots in the project VaMEx3.

[koide3]

@aserbremen
I'm so happy to see that small_gicp has been integrated into mrg_slam! Thanks a lot for letting me know that! The setup looks good; small_gicp can be used as just a drop-in replacement for fast_gicp ;)