search

HKUST-Aerial-Robotics / VINS-Mono

A Robust and Versatile Monocular Visual-Inertial State Estimator
GNU General Public License v3.0
4.95k stars 2.09k forks source link

Maybe a bug. visualInitialAlign() frame_i->second.R * x.segment<3>(kv * 3) out of range #68

Open Zekegao opened 7 years ago

Zekegao commented 7 years ago

Hi, I found in the visualInitialAlign(), maybe there is a bug that Vs[kv] = frame_i->second.R x.segment<3>(kv 3); Run it, kv could be 4, but x.segment<3> just 12 elements, so it will out of range. But, I cannot confirm it. Because I run it without ROS, so maybe my modification led to this bug. My log as below 2017-08-07 14-08-44

Zekegao commented 7 years ago

So, I add a condition, now I can run it. But, I don t know if this condition could affect all system.

if (kv == x.size()/3)
                break;
Vs[kv] = frame_i->second.R * x.segment<3>(kv * 3);
qintonguav commented 7 years ago

I have checked it on my code. x.size always large than kv 3. The extrem case is that x size is 36 kv 0 - 10 x.size = 11 3 (velocity) + 2 (gravity) + 1 (scale) image

Anyway, add your condition to keep safe.

Zekegao commented 7 years ago

@qintony Thx for your reply. I also found the Camera Matrix K in lots of place are not be set. it's usually 1 0 0 , 0 1 0 , 0 0 1. Should I change it to myself camera parameters? Actually, when I changed, the Ceres Solver will not convergence, and the final_cost also bigger than 5e-03, so the the GlobalSFM::construct will return false. If I need not to change the the camera parameters, why you didn't use the parameters by calibrate to calculate the cv::recoverPose and solvePnPRansac? Is there some reason? Best, zeke

qintonguav commented 7 years ago

You don't need to do that change. Because in the feature_tracker node, the features are already undistorted by K and D, which is normalized points.

If you think 5e-03 is too strict, you can release it.

Zekegao commented 7 years ago

@qintony Got it, Thank you very much! Another question.... After complete Initialization, the jacobian_speedbias_i.maxCoeff() will very very big, and then display Assertion `fabs(jacobian_speedbias_i.maxCoeff()) < 1e8' failed. I have checked, so I found the value of all_image_frame.second.R is big, like
2000.58 -973.176 -9.6511 229.516 -2.339 -5.24738 -15.8506 9.85859 0 I dont know why, I check the calib_ric is seem like normal. relative_R in relativePose is also correct. Anyway, thank you very much. Best, zeke