Image processing with known camera locations

[nych....@gmail.com]

I need to process images from a stereo rig with fixed and known positions of 20 cameras.

How should I specify camera locations in COLMAP for processing?

[Johannes Schönberger]

You need to get a sparse or dense reconstruction?

[Andriy Nych]

Ideally, I need to get some sort of "panoramic depth map", so dense reconstruction seems to be the right thing for me.

Practically, anything that may bring me closer to that will be also OK.

I'm playing with COLMAP just for a couple of days and I am not sure if I understand its ideology correctly.

I tried to change Q(WXYZ) and T(XYZ) in database widget after the features were indentified and matched (right before sparse reconstruction) but they did not seem to have any effect (or I did something wrong).

By the way, could you please point me to some description of the math behind projecting points from real world to camera image using quaternions and translation vector ( Q(WXYZ) and T(XYZ) )?

Thanks,

Andriy

[Ce Be]

Hi,

the idea of COLMAP and similar implementations is focused on unknown camera paramters, so a lot of work is put into estimating these. I do not know if COLMAP can use these as prior input.
Estimating the depth with known parameters is not hard compared to the usual SfM process. Have you tried a reconstruction with COLMAP without adding the camera positions? This may acutally work dependent on the images.

How much do you know about your cameras and how exact is this knowledge?
Do you know exact positions as well as internal camera parameters?

A good illustration about the projection from world to camera can be found in the OpenCV documentation:

http://docs.opencv.org/2.4.13.2/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html

Not sure about other online resources, but there is a good book about the topic:

"Multiple View Geometry in Computer Vision" by Richard Hartley and Andrew Zisserman:

https://books.google.de/books/about/Multiple_View_Geometry_in_Computer_Visio.html?id=si3R3Pfa98QC&redir_esc=y

Seems there are some useful pages in the google books preview already.
best regards,
Carsten

[undermin...@gmail.com]

I know I asked this question on github but I realized I put it in the wrong place and there already seems to be a thread started so I thought I'd ask here.

What is the outcome of this? My basic issue is I need my meshes to not move one bit. I need them to have the same r xyz and t xyz every time. I have a camera rig that doesn't move. Is there a way for me to use the estimations from one mesh and then reuse them for another. It helps in my automated cleanup process that and I suspect speed up calculating the meshes.