Full edge remotion in optimization in frame

[hkpolyu-uav]

Full edge remotion in optimization in frame

[Ttoto]

I have never come across this situation before. I think there may be sth wrong with your g2o installation.
Have you tried the EuRoC dataset or our provided rosbag (named:melab_sn943222072828.bag)?

[zouyajing]

I reproduced the error before.

The pose are solved twice in the frame optimization.
If g2o provides very bad pose estimation in the first solver, it will delete all the edges and then start the second solver, which will initialize an empty graph.

For me, it happened when optical flow tracking is too bad and feature matching is totally wrong. My solution is to skip the frame if the quality of optical flow tracking is bad.

[rocomanmelli]

Exactly! That was what I thought. However, I have done some tests monitoring the average error reported in err variable by cv::calcOpticalFlowPyrLK in lkorb_tracking.cpp and I did not find anything strange till it just throws the g2o error. I have also saved the images with the landmarks on them and nothing seems to change drastically. And there is not anything abnormal in the differences between pnp_inlier_cnt, F_inlier_cnt and of_inlier_cnt.

[Ttoto]

Exactly! That was what I thought. However, I have done some tests monitoring the average error reported in err variable by cv::calcOpticalFlowPyrLK in lkorb_tracking.cpp and I did not find anything strange till it just throws the g2o error. I have also saved the images with the landmarks on them and nothing seems to change drastically. And there is not anything abnormal in the differences between pnp_inlier_cnt, F_inlier_cnt and of_inlier_cnt.

Hello, @rocomanmelli we have revise the inframe optimization workflow. The bug has been fixed. #6

[rocomanmelli]

Great! Now the system runs the entire rosbags. Thank you.