How to determine the global scale of camera poses?

[thbupt]

Great work! The model requires camera poses, does the translation vector need to have global scale, i.e., meters? For example, in the dataset RealEstate10K, it seems the translation vector has no global scale.

[haofeixu]

Hi, the translation doesn't need to have metric scale. For example, they could come from colmap. In this case, our prediction will be not metric as well, but the scale is consistent with the camera translation. To find the metric scale in such cases, additional information might be necessary. For example, if the size of an object could be known from the scene. If the input camera pose is metric, then our prediction will be metric as well.