the camera's trajectory

[voladorfly]

When the handheld camera movement, the camera's trajectory is easy to stick together.Then the rotating in one spot.How should solve?

[AlejandroSilvestri]

You can't. ORB-SLAM2 can't map on pure rotations. You must translate while rotating.

[voladorfly]

I translate while rotating.But the key frames are also easy to stick together。

[AlejandroSilvestri]

If you are monitoring orb-slam "live", you can see the number of tracked points, don't let them go to less than 30, because there is where orb-slam looses track.

You can translate laterally in order to "make" more keyframes with parallax, what orb-slam needs to triangulate new points.

And, if you move slower, you can give enough time to local mapping so it can finish processing new points and add them to map: you can see them popping all toghether in groups.

[hg-wang2015]

I got the same problem, but when i use the Kinect1 sensor, this problem never appear. So i think it caused by the Xtion sensor not the ORB_SLAM2 program, the probable reason maybe is the incorrect launch method for the Xtion sensor, making it can not send correct data to ORB node. BUT, so far, i can not solve this problem. If you did, please tell me. Thanks!

[Morpheusai]

Hi, i'm using a monocular camera. I want to get trajectory of every frame. ORB_SLAM and ORB_SLAM2 don't support out pose of each frame, but only the keyFrame. Please tell me why? thanks a lot.

[AlejandroSilvestri]

@Morpheusai

You can get the actual frame pose from

Tracking.mCurrentFrame.mTcw

I believe in main you can get

SLAM.mpTracker->mCurrentFrame.mTcw

[ichlubna]

@AlejandroSilvestri thank you for the information. So do I understand correctly that pure camera tilt and pan movements are not detectable with ORB_SLAM2 at all?

[ichlubna]

@AlejandroSilvestri Also thanks for the idea with the current pose but it doesn't seem to be working. The mTcw is empty and basically contains the same thing as the result of TrackMonocular(). This code:

std::cerr << SLAM.TrackMonocular(im,tframe) << std::endl;
std::cerr << SLAM.getTracker()->mCurrentFrame.mTcw << std::endl; //the same what you suggested, just made a getter

Prints the same result, providing pose matrix only for keyframes. Related to #436

[AlejandroSilvestri]

@ichlubna,

TrackMonocular() returns mTcw, they are the same thing, they are the Tcw transform matrix. System must be tracking or mapping, meaning it must have been initialized and not lost. Then Tcw will be a valid 4x4 euclidean transform. Remember Tcw is the world pose in the camera reference. You need its inverse Twc, the camera pose in the world reference.

[AlejandroSilvestri]

@AlejandroSilvestri thank you for the information. So do I understand correctly that pure camera tilt and pan movements are not detectable with ORB_SLAM2 at all?

All movements are detectable in tracking mode, provided you stay in map. While mapping new areas ORB-SLAM2, like any monocular SLAM, can't map on pure rotations pan and tilt. These systems need a camera displacement in order to get enough parallax to triangulate new mappoints. So you can rotate while translating.

[ichlubna]

@AlejandroSilvestri I see, thanks for the answers. The problem about the matrix is that the Tcw is empty for most of the frames. So what you're saying is that the matrix won't be available for the first run? My use-case is having an unknown video sequence where I need to get camera positions while streaming new frames. Is ORB_SLAM2 able to do so?

[AlejandroSilvestri]

@ichlubna ,

That's right. The main problem with offline video sequence is the visual SLAM initialization. ORB-SLAM2 doesn't initialize on any random sequence, and some camera moves are known to be better for initialization.

[ichlubna]

@AlejandroSilvestri Many thanks for the clarification!

[ichlubna]

@ichlubna ,

That's right. The main problem with offline video sequence is the visual SLAM initialization. ORB-SLAM2 doesn't initialize on any random sequence, and some camera moves are known to be better for initialization.

I'd like to ask one more thing if you don't mind. Is it possible to retrieve the previous camera positions after the initialization? Or do I have to send the images in again?

[AlejandroSilvestri]

@ichlubna

No, before initialization no camera pose was computed.

I use to reverse the video to process those images in the map and get their poses.

[ichlubna]

@AlejandroSilvestri Thank you very much!