search

KumarRobotics / msckf_vio

Robust Stereo Visual Inertial Odometry for Fast Autonomous Flight
Other
1.66k stars 592 forks source link

Doubt about cam1_points initialization #106

Closed rocomanmelli closed 3 years ago

rocomanmelli commented 3 years ago

Hi! I have been seeing the code and got a doubt about this if. As far as I understand, trying to initialize the features of the right camera in that way would only make sense if the 3D points belonged to the same plane (in space) which should not be the case in general. Is it just an approximated guess or I am missing something? Thanks in advance.

ke-sun commented 3 years ago

Not sure about the "the same plane" way of thinking. Probably it's easier to think in the following way.

What's in the condition is a way to general initial guesses for matches (which will be later refined through tracking). The guess is generated by simply projecting the pixel coordinates from the left frame to the right frame. If the rotation between left and right cameras is identity, and the calibrations parameters of both cameras are the same, cam1_points would be just a copy cam0_points. It may not be the best way to generate initial guess, but should be sufficient most times.

https://github.com/KumarRobotics/msckf_vio/blob/e3a39a95b57da376924aa22b46637b60f68f504c/src/image_processor.cpp#L616-L626

rocomanmelli commented 3 years ago

Thank you for the answer. I understand. When I asked I was thinking about a case like this: issue_s_msckf For P, it would not be a good initial guess, right?

rocomanmelli commented 3 years ago

Anyway, the condition is not related to belonging to the same plane in space or not, actually. I was wrong about that. It is a matter of stereo disparity I guess.

ke-sun commented 3 years ago

I see. In that sense, the above initialization assumes all 3-d points are infinitely far away from the image plane.