A problem about the camera scale xyz

[xinfeixiang1]

Thanks for your great work.

However, in our practical tests (where the input is a video), we found that the camera's position in space was normalized to a range with a diameter of 6, regardless of whether the camera was moving closer to or longer in space. This does not align with the actual scale of the camera's position. Could you please advise on how to resolve this issue?

Thank you very much.

[jytime]

Hi @xinfeixiang1 ,

Are you aiming to reconstruct the scene at a real-world scale? Unfortunately, this isn’t feasible with structure from motion (SfM), as it inherently produces scale-agnostic reconstruction. If accurate scale is crucial, you could introduce additional references, such as known size of an object, to calibrate the SfM output to real-world dimensions.

Regarding camera motion, it can only be understood relatively through SfM. For example, you can determine that one camera position is 10% closer to an object compared to another, but you won’t be able to quantify the exact distance in meters.

[xinfeixiang1]

Hi @jytime
Thank you for your response so quick. I understand that SfM can only provide relative scale without additional references. My current issue is about recovering the camera motion from a video, specifically using the extrinsics_opencv parameter in vggsfm/runners/runner.py. However, I have noticed that regardless of whether the video's movement is fast or slow, the radius of the camera's external motion obtained through extrinsics_opencv remains approximately the same. Could you kindly advise on how to address this issue? Thank you.

[jytime]

I am still confused by what you may want.. But if you hope to stop the normalization of the cameras, just skip this:

vggsfm/vggsfm/utils/triangulation.py

Line 1217 in 4eb6fae

reconstruction.normalize(5.0, 0.1, 0.9, True)