search

raulmur / ORB_SLAM2

Real-Time SLAM for Monocular, Stereo and RGB-D Cameras, with Loop Detection and Relocalization Capabilities
Other
9.26k stars 4.69k forks source link

Monocular SLAM Results on EuRoC dataset #741

Open abhilashjain7 opened 5 years ago

abhilashjain7 commented 5 years ago

Hi! I cannot find any results for monocular ATE on EuRoC dataset. The ORB-SLAM2 paper gives the results for stereo SLAM on EuRoC. If someone has also evaluated the code on EuRoC (monocular), could you please validate my results?

Stream ATE - RMSE
MH_01_easy 0.037298
MH_02_easy 0.031539
MH_03_medium 0.03853
MH_04_difficult 0.048295
MH_05_difficult 0.04384
V1_01_easy 0.088099
V1_02_medium 0.064248
V1_03_difficult 0.767821
V2_01_easy 0.060076
V2_02_medium 0.0532
V2_03_difficult 0.164655
thien94 commented 5 years ago

This paper evaluated ORB-SLAM2 on most popular datasets for (EuRoC, KITTI, TUM). Here are their results on EuRoC: Picture1

abhilashjain7 commented 5 years ago

Thank you! However, after going through the paper, I realized that these results are also on the stereo SLAM. I wanted to see if there's any paper which gives the results on monocular SLAM.

thien94 commented 5 years ago

Have you checked out Raul Mur-Artal's paper "Visual-Inertial Monocular SLAM with Map Reuse" (see table I, right-most 2 columns)?

For other methods (ROVIO, MSCKF, VINS-Mono etc.), ATE results are readily available in this paper.

I am curious about how you get the metric scale for monocular ORB-SLAM2 in your results. Did you use the ground truth to scale up, or are you using your own method to calculate true scale from just one camera? If it is the latter, I look forward to seeing your findings.

abhilashjain7 commented 5 years ago

Thanks for linking these two papers. I had somehow missed these two when I was looking for the results. To get the right scale, I simply used the modified ATE script given here, which is a modification of the evaluation script given with the TUM RGB-D dataset. This script automatically aligns the scale of the estimated trajectory to the ground truth for monocular SLAM.

sazak1 commented 3 years ago

Hello @abhilashjain7, I ran the ORB-SLAM code with EuRoC using both monocular and stereo vision. But I couldn't get overlapping trajectory and ground-truth with any type of camera. Could you draw overlapping trajectories with the ground-truth data? Actually I do not understand to adjust scale factor or anything else manually unless we work with monocular. So, after you create the KeyFrameTrajectory.txt file what kind of a procedure did you follow?