tum-vision / lsd_slam

LSD-SLAM
GNU General Public License v3.0
2.59k stars 1.23k forks source link

correct parameter choice #168

Open sharder-dev opened 8 years ago

sharder-dev commented 8 years ago

I am trying to use LSD-SLAM with the camera provided in the 'Parrot Bebop Drone' but the results aren't as good as I hoped they would be. The 1080p@30FPS video stream from this specific camera is cropped to 640x480 and calibrated (I did this multiple times just to make sure). I know how to create the pointcloud from the keyframes and a few test-videos resulted in usefull points (but with lots of outliers). Even though I tried multiple different cameras and parameters modified according to the notes, it never looks good enough (e.g. the pose estimation in longer videos became completely wrong or tracking was lost after some rotations). I assume the problem lies in the cameras I tried (small fov of ~60-80) or my calibrations are still bad. Could someone push me in the right direction to find the correct parameters/solution I need to make it work with the drone camera?

As an example I get the following result using the bebop drone camera: noise The initialization: noise_2

To create the pointcloud, I use the method described in lsd-slam viewer. Parameters used in the example above:

EDIT: added an example

klekkala commented 7 years ago

@Seryogah .. Any updates on this issue? And BTW... FoV for Parrot Bebop's camera is pretty high.. 180 or so.. Did you check that?

kenisvery commented 7 years ago

@Seryogah I'm also doing relevance project. Could you share the package on Github?