search

facebookresearch / consistent_depth

We estimate dense, flicker-free, geometrically consistent depth from monocular video, for example hand-held cell phone video.
MIT License
1.61k stars 236 forks source link

Scale calibration #47

Open jaykar21 opened 3 years ago

jaykar21 commented 3 years ago

In section 4 under scale calibration to align the colmap scale with the network, the global scale adjustment factor is multiplied with the translation vector as shown in eq. 3. In the implementation the translation vector is divided by the scale factor. Is there a reason for this difference or am I missing something here.

Qjizhi commented 1 year ago

In section 4 under scale calibration to align the colmap scale with the network, the global scale adjustment factor is multiplied with the translation vector as shown in eq. 3. In the implementation the translation vector is divided by the scale factor. Is there a reason for this difference or am I missing something here.

The reason why there is a scale is because colamp and not get the absolute 3d information with only a sequence of images, so the semi-depth from colmap has a scale with the real depth, which is inferred from learned-based model. So does translation