Hi, I am new to SLAM, is it possible to provide more detail instructions on conducting the comparison experiment in your repo? Or can you tell me where can I find those instructions, thanks
@@Eungchang_Mason_Lee thanks for replying, I have gone to the repo and the fastlio's instructions to simulate using rosbag. But how do I compare two LIO algorithms like what you have done in these videos? I want to see which algorithm work best for my dataset, do I need to record it with rosbag first
@@Eungchang_Mason_Lee 고맙습니다. 근데..roslaunch 까지는 되는데, mid360이 연결되는 시점부터는 process has died가 떠버립니다. imu Initials 하다가 No point!!! 이렇게요.죄송합니다,, 너무 초보라.. 방법이 없을까요?
What's your comment on that ?? I really like ImMesh for it's mesh generation ability but from your video it seems like the SLAM method (probably Fast LIO) they are using is not that good. What's your suggestion on that ? Should I use a different SLAM with ImMesh instead of FastLIO ?? I will be using this for quadrupeds exploring multiple levels, so the SLAM needs to be drift free.
There are 2 more additional questions. 1. Did you see any performance improvement when you used Ouster over livox ?? I plan to use Ouster for now but may probably switch to Livox later. So, the SLAM methods which doesn't support Livox may unfortunately out of scope for me. 2. Is it able to generate colored mesh as well ?? I mean am I able to see that colored mesh inside Rviz realtime ??
Inmesh didn't opensurce the colour mapping module by performance issues . I guess ADA-Lio is based in some other stack and it is not open sourced to don't confront in Licenses or keep it in personal or academic use to fit to a License which is fair and Legit ,both use and don't open it. How perform every stack , there is so many variables that can make one perform better than other that they are unlimited, just test every of them in your application. For sure Fast-Lio is in the top 3 and possibly the top one in reconstruction quality. Only some specific apps compete in specific environments. More features, as mesh needs more computing power, so need more powerful computer. Say a more featured software perform worse becasue use the same computer than other have not many sense, as resources affects to the mapping, except if the comparison is about resources requeriments and there, no Faslio, no adalio no inmesh will win for sure. For sure as this video is useful for get some ideas, but you cannot judge the reliability of the software im this way, and I'm sure is not the target of this video, which I'm sure is informative and no comparive.
@@Flamenawer Well, changing a license of a work is not that difficult, you can always ask the author to negotiate that and he/they can do that if a company or entity wants to use it for their professional work. Also, I think comparison is good since that's how people do work / progress in academia since if there is not something left for improvement, that field is dead. You can see this work: ua-cam.com/video/b2lBQ-q5rOE/v-deo.htmlsi=gIpxAnXVKKojyzSi You can see how this paper tries to compare with the SOTA methods since an academician's paper won't get accepted in conference / journal if you don't do a certain sort of benchmarking against SOTA. Whether he is not showing any comparison or not that I am not sure, I personally feel that his videos are good for initial comparison of different methods which is very important. Because if you are not working on something which has a good performance (in SLAM, it's pose accuracy or APE of the robot), then you can not iterate over that and make it better. So, taking a good performing algorithm as a baseline and working on it is a good way to improve performance of a system in academia and publish new papers. A Good way SLAM methods do that is by performing ablation study which helps to figure out how a SLAM system performs under difficult conditions. It's mostly written in the paper. Since he has sufficiently tested his system as well as this one (ImMesh), I believe he will be able to tell the difference he faced while running them.
@@ArghyaChatterjeeJony ImMesh uses VoxelMap(github.com/hku-mars/VoxelMap) as its odometry, referring the paper. VoxelMap estimates probabilistic planes and uses them to estimate odometry. From my experience, this method shows more robust result in wide open area but drifts worse than Fast-LIO in narrow environments. I usually prefer Fast-LIO than VoxelMap (or PV-LIO) considering the computational usage but still robust and accurate performance. If you want drift free results, then you should consider loop-closing added ones. How about try my other works like FAST-LIO-SAM-QN? For additional questions, 1. if you look into the config files of Fast-LIO or its variants, they are downsampling the points a lot with voxelization and point_filter_num. Ouster scans even more points than Livox but you can just tune the numbers of points you use in the config files. So what you have to consider is the FoV and the precision of the sensor. For me, I do love both of them and use them depending on the environments and tasks of the robots. 2. With ImMesh, Yes. There is a such function.
@Flamenawer @ArghyaChatterjeeJony The purpose of my video is mainly to provide the first glance of the state-of-the-art methods' performances in general environments including where they perform well and where they cannot.
![[SLAM] Ada-LIO vs PV-LIO in real-world and Gazebo (test#16) - Ouster, Livox mid-360](http://i.ytimg.com/vi/0sVkY66pOpI/mqdefault.jpg)
![[SLAM] Ada-LIO vs PV-LIO in real-world and Gazebo (test#16) - Ouster, Livox mid-360](/img/tr.png)
Hi, I am new to SLAM, is it possible to provide more detail instructions on conducting the comparison experiment in your repo? Or can you tell me where can I find those instructions, thanks
Hi. Check out here - github.com/engcang/slam-application
@@Eungchang_Mason_Lee thanks for replying, I have gone to the repo and the fastlio's instructions to simulate using rosbag. But how do I compare two LIO algorithms like what you have done in these videos? I want to see which algorithm work best for my dataset, do I need to record it with rosbag first
확실히 ImMesh는 CPU 상당히 많이 잡아먹긴 하더라고요.. 혹시 Ada LIO는 open될 계획이 있을까요...?
Ada LIO는 제가 1저자가 아니라서... 그리고 기술이전도 해서 아마 open은 몇 년 뒤에나 가능하지 않을까 싶습니다...
@@Eungchang_Mason_Lee 아하.. 그런 배경이 있었군요! 답변 감사드립니다 :)
@@donghachung 넵 Fast LIO에 동작 모듈?이 붙은거라 Fast LIO의 조금 개선 버전이라고 보시면 됩니닷
@@Eungchang_Mason_Lee 네 ㅎㅎ 직접 구현해봐야겠네요! 연휴 잘 보내시고 조만간 또 커피 한잔 해요 :)
안녕하세요 실험영상은 잘 보고 있습니다, 최근에 mid360으로 test 해보려고 합니다. 근데 immesh file 안에 mid360 launch file이 없더라고요.. mid 360 구성은 어떻게 해야 하나요?
avia launch파일로 사용하시면 돼요~
@@Eungchang_Mason_Lee 고맙습니다. 근데..roslaunch 까지는 되는데, mid360이 연결되는 시점부터는 process has died가 떠버립니다. imu Initials 하다가 No point!!! 이렇게요.죄송합니다,, 너무 초보라.. 방법이 없을까요?
What's your comment on that ?? I really like ImMesh for it's mesh generation ability but from your video it seems like the SLAM method (probably Fast LIO) they are using is not that good.
What's your suggestion on that ? Should I use a different SLAM with ImMesh instead of FastLIO ?? I will be using this for quadrupeds exploring multiple levels, so the SLAM needs to be drift free.
There are 2 more additional questions.
1. Did you see any performance improvement when you used Ouster over livox ?? I plan to use Ouster for now but may probably switch to Livox later. So, the SLAM methods which doesn't support Livox may unfortunately out of scope for me.
2. Is it able to generate colored mesh as well ?? I mean am I able to see that colored mesh inside Rviz realtime ??
Inmesh didn't opensurce the colour mapping module by performance issues .
I guess ADA-Lio is based in some other stack and it is not open sourced to don't confront in Licenses or keep it in personal or academic use to fit to a License which is fair and Legit ,both use and don't open it.
How perform every stack , there is so many variables that can make one perform better than other that they are unlimited, just test every of them in your application.
For sure Fast-Lio is in the top 3 and possibly the top one in reconstruction quality. Only some specific apps compete in specific environments. More features, as mesh needs more computing power, so need more powerful computer.
Say a more featured software perform worse becasue use the same computer than other have not many sense, as resources affects to the mapping, except if the comparison is about resources requeriments and there, no Faslio, no adalio no inmesh will win for sure.
For sure as this video is useful for get some ideas, but you cannot judge the reliability of the software im this way, and I'm sure is not the target of this video, which I'm sure is informative and no comparive.
@@Flamenawer Well, changing a license of a work is not that difficult, you can always ask the author to negotiate that and he/they can do that if a company or entity wants to use it for their professional work. Also, I think comparison is good since that's how people do work / progress in academia since if there is not something left for improvement, that field is dead.
You can see this work: ua-cam.com/video/b2lBQ-q5rOE/v-deo.htmlsi=gIpxAnXVKKojyzSi
You can see how this paper tries to compare with the SOTA methods since an academician's paper won't get accepted in conference / journal if you don't do a certain sort of benchmarking against SOTA. Whether he is not showing any comparison or not that I am not sure, I personally feel that his videos are good for initial comparison of different methods which is very important. Because if you are not working on something which has a good performance (in SLAM, it's pose accuracy or APE of the robot), then you can not iterate over that and make it better. So, taking a good performing algorithm as a baseline and working on it is a good way to improve performance of a system in academia and publish new papers. A Good way SLAM methods do that is by performing ablation study which helps to figure out how a SLAM system performs under difficult conditions. It's mostly written in the paper. Since he has sufficiently tested his system as well as this one (ImMesh), I believe he will be able to tell the difference he faced while running them.
@@ArghyaChatterjeeJony ImMesh uses VoxelMap(github.com/hku-mars/VoxelMap) as its odometry, referring the paper. VoxelMap estimates probabilistic planes and uses them to estimate odometry. From my experience, this method shows more robust result in wide open area but drifts worse than Fast-LIO in narrow environments.
I usually prefer Fast-LIO than VoxelMap (or PV-LIO) considering the computational usage but still robust and accurate performance. If you want drift free results, then you should consider loop-closing added ones. How about try my other works like FAST-LIO-SAM-QN?
For additional questions,
1. if you look into the config files of Fast-LIO or its variants, they are downsampling the points a lot with voxelization and point_filter_num. Ouster scans even more points than Livox but you can just tune the numbers of points you use in the config files. So what you have to consider is the FoV and the precision of the sensor. For me, I do love both of them and use them depending on the environments and tasks of the robots.
2. With ImMesh, Yes. There is a such function.
@Flamenawer @ArghyaChatterjeeJony The purpose of my video is mainly to provide the first glance of the state-of-the-art methods' performances in general environments including where they perform well and where they cannot.