Basic Question: Obtaining Camera Extrinsics

[joanac...@gmail.com]

Hi,

I apologize for my basic question, but I am using Sfm for the first time to get camera calibration parameters and something is not right.

I was hoping that someone could help.

COLMAP does, what I think it is, a great job in reconstructing the scene and the position of the cameras are where they suppose to be.

However when I use the calibration parameters in opencv for point triangulation, I am getting a very large reprojection error.

I am wondering if I am calculating something wrong.

This is COLMAP output:

fx=879.544

fy=908.766

cx=959.03

cy=535.747

k1=-0.0859365

k2=0.245244

k3=-0.199789

k4=0.0703465

QW=-0.0963985

QX=-0.0165644

QY=0.960941

QZ=0.258893

TX=0.85698

TY=-2.57258

TZ=4.47233

Which results is the following extrinsic matrix:

[-0.98086638  0.01807897 -0.19384336 -0.15993675

 -0.08174862  0.86540007  0.49436823 -0.02872564

 0.17668973  0.50075536 -0.84736397  5.227613]

The last column of the matrix is given by: -R^t * T

Also, my distortion vector according to opencv format is: [-0.0859365 0.245244 0 0 -0.199789 0.0703465]

Is this correct?

Is it the case that the problem might be in the intrinsic parameters of the camera, which were estimated using the model + bundle adjustment?

Should I calibrate the cameras individually (to get the intrinsics) and then re-run the SfM?

(Note: the images were captured using gopro hero3 black and I used opencv_fisheye camera model for the reconstruction)

Thank you very much

[Johannes Schönberger]

If you use the Python script here to read the model (https://github.com/colmap/colmap/blob/dev/scripts/python/read_model.py), then you can get the projection center of the camera as ``np.dot(-image.qvec2rotmat().T, image.tvec)``