Fatal Error (variance is not finite)

[g.bartoli]

Hi Mathieu,
when using RTAB-Map, sometimes after some minutes working, the program gives this error:

[FATAL] (2015-08-31 09:13:52.757) Link.h:110::setVariance() Condition (uIsFinite(rotVariance) && rotVariance>0) not met!

*******
FATAL message occurred! Application will now exit.

*******
Killed

And there is no way to restore it... do you know why or in which circumstances can it occur? If it can be useful, the error occured mainly when the sensor was not moving (i.e. sitting still on a table).

Thanks!

[matlabbe]

Hi,
Which node is giving this assert? Are you using the standalone version? or ROS? Normally, FATAL errors are here to debug when the values are incorrect. For some reasons, the Link::setVariance() is called with null variance.

Which version are you using? I'll try to reproduce the problem.

cheers

[g.bartoli]

I started RTAB-Map standalone from the command line with "rtabmap", so I think the inner node causing the error was the main rtabmap, correct?
I installed it normally with "sudo apt-get install ros-indigo-rtabmap", it's 0.10.4 version.

I changed some parameters starting from your config.ini provided here, like using FAST+FREAK for Visual Word, decreased the acquisition and detection frequency, but not much more... Since then, it's not happened again, I'm trying to reproduce it and if it will represent again, I will send to you the config file and log file, ok?

[matlabbe]

Hi,

You can also set the logger's level to Debug or Info for more information (default Warning). The setting is in Preferences->General Settings (GUI)->Logging panel. And yeah, the config file used and the log would be great to debug.

thx,
Mathieu

[g.bartoli]

Hi Mathieu,
I have some more information about this fatal error.

It occurred many more times on RTABMAP standalone 0.10.4 (latest apt-get repository package), both during SLAM and post-processing (without ICP). I think it is related to the loop constraint procedure, in particular to the RANSAC iterations using util3d::transformFromXYZCorrespondence, this is the debug log happening just before it:

[DEBUG] (2015-10-15 11:23:28.648) Memory.cpp:778::addSignatureToStm() time = 0,000121s
[DEBUG] (2015-10-15 11:23:28.648) Memory.cpp:606::update() time rehearsal=0,181913 ms
[DEBUG] (2015-10-15 11:23:28.648) Memory.cpp:673::update() totalTimer = 0,038850s
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:70::transformFromXYZCorrespondences() iterations=200 inlierThreshold=0,020000
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 197/199, with an error threshold of 0,020000.
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,014632 (variance=0,000024) on iteration 0 out of 10.
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 199/191, with an error threshold of 0,014632.
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,009591 (variance=0,000010) on iteration 1 out of 10.
[DEBUG] (2015-10-15 11:23:28.648) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 191/176, with an error threshold of 0,009591.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,005935 (variance=0,000004) on iteration 2 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 176/152, with an error threshold of 0,005935.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,001354 (variance=0,000000) on iteration 3 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 152/94, with an error threshold of 0,001354.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000819 (variance=0,000000) on iteration 4 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 94/86, with an error threshold of 0,000819.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000101 (variance=0,000000) on iteration 5 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 86/78, with an error threshold of 0,000101.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000121 (variance=0,000000) on iteration 6 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 78/79, with an error threshold of 0,000121.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000007 (variance=0,000000) on iteration 7 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 79/76, with an error threshold of 0,000007.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000006 (variance=0,000000) on iteration 8 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:119::transformFromXYZCorrespondences() RANSAC refineModel: Number of inliers found (before/after): 76/78, with an error threshold of 0,000006.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:136::transformFromXYZCorrespondences() RANSAC refineModel: New estimated error threshold: 0,000000 (variance=0,000000) on iteration 9 out of 10.
[DEBUG] (2015-10-15 11:23:28.649) util3d_registration.cpp:204::transformFromXYZCorrespondences() RANSAC inliers=76/315 tf=xyz=0,000000,0,000000,-0,000000 rpy=-0,000000,-0,000000,-0,000000
[DEBUG] (2015-10-15 11:23:28.649) Memory.cpp:2208::computeVisualTransform() transform=xyz=0,000000,0,000000,-0,000000 rpy=-0,000000,-0,000000,-0,000000
[ INFO] (2015-10-15 11:23:28.649) MainWindow.cpp:3455::postProcessing() Added new loop closure between 124 and 122.
[FATAL] (2015-10-15 11:23:28.649) Link.h:110::setVariance() Condition (uIsFinite(rotVariance) && rotVariance>0) not met!

As you can see, RANSAC reduces both error and variance almost to 0 obtaining a final null transform. Probably the following rotVariance value is 0 too or is computed via some division by zero invalidating the mandatory condition.

This is my current config file, apart from some parameters tweaking, I enabled "Force 2D transform" in Odometry and Loop Closure Constraint, maybe this can be an hint.

UPDATE: actually, this error occurs even when the forced 3-DoF constraints are disabled in "Graph Optimization", "Loop Closure Constraint" and "Odometry".

Let me know if you need more informations.

Have a nice day,
Guido

[matlabbe]

Are you in a simulated environment or replaying exactly the same data? A null variance means that all the correspondences are exactly one over the others, which rarely happens in real situations. This is indeed a bug, which is fixed in latest versions. When the 0 variance is detected, it is set to a small variance epsilon (0.00001).

cheers

[g.bartoli]

I'm teleoperating our robot in order to map the indoor office. I'm not using bag files, just repeating real experiments in the same enviroments, the actual paths are similar but not identical.

In fact that error happens after the robot is stopped and remains in still position for some time, so maybe the stream of very similar images occasionally produces features that are coincident, even if I still believe it is pretty rare... however, thanks for fixing this in the recent commit, I will compile the latest version on the robot software, too!

Thanks,
Guido