FOVIS odometry produces NaNs
Good day everyone.
I've tested FOVIS library with Juergen Sturm's datasets and discovered that sometimes (approximately every 500 frames) fovis::VisualOdometry may produce some NaNs in the estimated transformation. The input datasets have no NaN; unknown depth is encoded by zeroes.
Here are the samples: http://vision.in.tum.de/data/datasets/rgbd-dataset. In particular I've used Pioneer and Desk 360.
Now I simply create a new instance of fovis::VisualOdometry every time I see NaN in output, but I don't like this kind of voodoo programming much.
Is it a flaw of the library itself, or maybe there is something I need to fix?
Thanks.