search

UZ-SLAMLab / ORB_SLAM3

ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM
GNU General Public License v3.0
6.68k stars 2.57k forks source link

The RGBD SLAM perform worse than ORB SLAM2 #208

Closed ZhouXiner closed 2 years ago

ZhouXiner commented 3 years ago

I try to run the RGBD type in TUM RGBD Dataset, but the performance is even worse than ORBSLAM2 in some sequence like fr1_desk2. Has anyone struggled the same problem like me?

ZhouXiner commented 3 years ago

How is it worse? Only changes in code compared to ORB-SLAM2 are : ORB_SLAM3 CV_LOAD_IMAGE And the difference in the Slam system is added imu support. I don't think they changed the way how no Imu systems work.

Thanks for reply. The APE is bigger in TUM RGBD dataset like fr1_desk2 and fr1_room. Actually, the loop closuing part also changed a lot. Maybe I should compare the preformance without loop.

melhashash commented 3 years ago

I am experiencing the same issue with the monocular pipeline; the accuracy is worse than ORB-SLAM2.

CSU-NXY commented 3 years ago

I've tested ORB-SLAM2 and ORB-SLAM3 on KITTI dataset in RGBD mode with my deep learning model predicted depth. It turns out that ORB-SLAM3 performs worse than ORB-SLAM2 and have scale drift issues.

CSU-NXY commented 3 years ago

Hi, @ZhouXiner @melhashash Try this solution #107. It seems the RGBD mode didn't set mbf value correctly.

ZhouXiner commented 3 years ago

Hi, @ZhouXiner @melhashash Try this solution #107. It seems the RGBD mode didn't set mbf value correctly.

Thanks, I will try it.

melhashash commented 3 years ago

Hi, @ZhouXiner @melhashash Try this solution #107. It seems the RGBD mode didn't set mbf value correctly.

Thanks for your suggestion, but I am talking about the monocular case.

thomCastillo commented 3 years ago

Could we add depth sensor error as parameter(i.e noise parameter in IMU) to optimization in RGB-D SLAM? If so, how can we add it?

UcefMountacer commented 3 years ago

@ZhouXiner can you give me the link to dataset and the command? I tried on freiburg desk dataset but it doesn't work for me.

jdtardos commented 2 years ago

Version v1.0 has corrected some bugs in RGB-D SLAM