Question about the fixed intrinsic you used in the KITTIDataset class about monodepth2 HOT 5 CLOSED

@wuyujack the post-rectification projection matrix has identity rotation matrix. So, projection matrix can be decomposed into intrinsics and extrinsics matrix, where the rotation component of the extrinsics matrix is identity.

In general, you need to do calibration to estimate intrinsics and extrinsics. Then you can rectify these matrices for the two views and compute projection matrices in a new space where the optical axes of the two cameras are parallel to each other (i.e. the rotation matrix is identity). This OpenCV function does the rectification:
https://docs.opencv.org/3.3.1/d9/d0c/group__calib3d.html#ga617b1685d4059c6040827800e72ad2b6

from monodepth2.

This makes me confusing. Based on what I know, the projection matrix should be composited by the camera intrinsic matrix and extrinsic matrix, but based on your reply it seems like you are using the projection matrix directly as the intrinsic matrix? Is it a default usage for the Kitti dataset? I am also verifying this since now I only obtain the camera intrinsic matrix based on the focal length and central points and I wonder whether I still need to get the extrinsic matrix which is composited by the translation vector and the rotation matrix.

from monodepth2.

All our code actually assumes a single intrinsics so you just need to change self.K
Just beware that the flip augmentation assumes the principal point to be in the centre of the image.

from monodepth2.

@mrharicot will be able to correct me if I'm wrong here, but I think the KITTI intrinsics you should be looking at are the post-rectification matrices. These are P_rect_02 and P_rect_03 in calib_cam_to_cam.txt. The maths using these matricies should work out:

P_rect_03: 7.215377e+02 0.000000e+00 6.095593e+02 -3.395242e+02 0.000000e+00 7.215377e+02 1.728540e+02 2.199936e+00 0.000000e+00 0.000000e+00 1.000000e+00 2.729905e-03

Then for example looking at f_x:

7.215377e+02 / 1242 = 0.58094....

...which is approximately what we have in self.K.

from monodepth2.

@wuyujack the post-rectification projection matrix has identity rotation matrix. So, projection matrix can be decomposed into intrinsics and extrinsics matrix, where the rotation component of the extrinsics matrix is identity.

In general, you need to do calibration to estimate intrinsics and extrinsics. Then you can rectify these matrices for the two views and compute projection matrices in a new space where the optical axes of the two cameras are parallel to each other (i.e. the rotation matrix is identity). This OpenCV function does the rectification:
https://docs.opencv.org/3.3.1/d9/d0c/group__calib3d.html#ga617b1685d4059c6040827800e72ad2b6

Thank you for your detailed reply and if the post-rectification projection matrix has identity rotation matrix then it makes sense to me right now.

By the way, I only use a single camera to obtain own color image sequence, so it is no need to do rectify and I can just use my camera intrinsic matrix to replace the one in your code (self.K), is it correct?

from monodepth2.