Is the camera extrinsic pose obtained with a scale, or is it like colmap, with no scale factor

[Ghul-huan]

Thank you very much for your open source. I would like to ask about the scale of external references： ① Is the camera extrinsic pose obtained with a scale, or is it like colmap, with no scale factor？ ② If the external reference has no scale, how can I restore the scale? Now there is an initial camera pose (external reference)?

[Ghul-huan]

Or Can we set the initial value for internal and external references？

[hxy-123]

Since we use COLMAP for coarse SfM and then refine it, the reconstructed SfM model is in arbitrary world coordinates without real metric scale. For your point (2), it is unfortunate that our system currently does not support the real-scale recovery. If you have a few images with known metric poses, I think maybe you can try to align the SfM results to the known poses.

[Ghul-huan]

The internal reference and pose (external reference, 3D point cloud coordinates) in the paper are obtained by minimizing the reprojection error. Is it feasible to set initial values for the internal reference and pose (external reference) in the code, and then optimize the internal reference and pose (external reference) through iteration to obtain the scaled point cloud coordinates?

[Ghul-huan]

设置内外参初值，在代码中迭代优化内外参，并求解3d点云

[hxy-123]

I'm wondering whether you have the initial poses for all images or only a subset of images. For the former case, you can perform triangulation (with fixed poses) in the coarse SfM and then refine the point cloud as well as poses in refinement phase.

[Ghul-huan]

（1）My scene is a self-driving scene, and all the images have an initial pose, but it is not accurate. （2）I'm not sure if my understanding is correct: I don't understand what you mentioned above. According to the paper, the initial Coarse SfM mainly produces initial matching pixel points, and does not use 3D points or poses (external parameters). The external parameters are calculated through iterative reprojection error, and according to the theory, the initial values of external parameters or even internal parameters should not be set in Geometry Refinement

[hxy-123]

The SfM does not need to set the initial values (both intrinsics and extrinsics), as shown in the paper and evaluations. If I'm not misunderstood, your requirement is to use the known rough poses with real-scale as the initial value to obtain SfM models with real-scale. If so, even if our system does not directly support this, I propose a solution (suggestion): "you can perform triangulation (with fixed poses) in the coarse SfM and then refine the point cloud together with poses in refinement phase."