TRACKING LOST for frame x (x = small number) and "cholesky failed!"

[tum-vision]

TRACKING LOST for frame x (x = small number) and "cholesky failed!"

[siims]

It has been a while... but anyway, I got back to it and I confirm that it was camera calibration problem. After I parsed my values to PTAM configuration format then things started to work magically. I also made a pull request: #34
Thank you very much for help :)

[JakobEngel]

sounds like it's diverging entirely - which shouldn't happen with careful motion and sufficiently structured / textured environment, even with a relatively unsuited camera.
Maybe your calibration is bad? If you send me a .bag of an example sequence, I can have a look at it.

[siims]

Here is bagfile (https://dl.dropboxusercontent.com/u/4148558/OpenCV_example_calib.cfg) and calibration (https://dl.dropboxusercontent.com/u/4148558/2014-10-17-12-23-19.bag).

[JakobEngel]

hmm, that bagfile contains a 10s video of a stationary camera... there is not much to reconstruct.

[siims]

Ok, I think I am doing something totally wrong. I tried out provided dataset from http://vision.in.tum.de/research/lsdslam and I am having the same trouble with the "Desk Sequence" bagfile.
I shortly describe what I am doing, maybe this helps you to understand where I am failing:

1. run lsd_slam "rosrun lsd_slam_core live_slam image:=/image_raw camera_info:=OpenCV_example_calib.cfg"
2. hold camera (30fps, rolling shutter, fov 75 deg angle) in my hand, press 'r' while moving camera in the same image plane (have tried sideways, vertical and circular motions).
   ... and with this depth map doesn't converge.

Something I leave undone? Do wrong?

[JakobEngel]

for the desk sequence you should use the camera calibration included tin the .bag, and not the OpenCV_example_calib.cfg. i.e., run
rosrun lsd_slam_core live_slam image:=/image_raw
rosbag play LSD_room.bag

[siims]

also now you can remove "[untested]" from README near OpenCV camera calibration file.