Questions about camera calibration

[zhangwenniu]

Hi, thanks for your good jobs.

Here are some questions about using AceZero as camera calibration framework.

1. Can this work be used to replace COLMAP to obtain the camera pose? Use the camera calibration results of this work for subsequent other 3D reconstruction work, such as for the training of 3D Gaussian Splatting.
2. If it can be used for camera calibration, are there any scripts that can convert the results output by this paper into the format of COLMAP, so that it can be used in other reconstruction methods in the 3D reconstruction community that are based on COLMAP calibration results?
3. Regarding running time and video memory usage. COLMAP calibrates 100 pictures with a resolution of 1024x768, which takes up approximately 2GB of GPU video memory, and the training time is approximately 10 to 30 minutes. If this paper is only used to obtain the camera pose, under the picture scale and resolution mentioned above, approximately how much video memory and training time overhead are required?

Thank you very much for your attention. I look forward to your reply at your convenience.

[ebrach]

Hi!

1) Yes. But please be aware of some restrictions of the current implementation. For example, ACE0 is limited to processing images with shared intrinsics and without significant image distortion.

2) No. The pose file format of ACE0 is roughly following COLMAP conventions but there are no direct conversion scripts available. We support conversion to Nerfstudio inputs if that helps.

3) The runtime depends on the spatial distribution of cameras in the scene - see some discussion on this in the paper supplement A. In general, it is unlikely that ACE0 is faster than COLMAP for 100 images. I think you will see benefits in speed if you reconstruct 1000 images and more. GPU memory consumption of ACE0 is discussed in the FAQ in the README here on Github.

Hope that helps, Eric