Closed hardikdava closed 5 months ago
@maturk got my answer.
Right yeah, so I solve for a per frame scale-and-shift for datasets with COLMAP SfM data (sparse point cloud basically). This aligns the monocular depth estimates with the poses estimated from COLMAP. This could be extended to work with arbitrary datasets as well, but currently only COLMAP formatted data is supported.
@maturk great work, I am curious about your method of aligning sfm depth and monocular estimated depths. In your method, it tries to find global scale and shift by optimizing it for every estimated depths. But since there is no any specific scale or shit present in monocular depths, does not it requires separate scale and shift for individual image rather than opt for global scale? It would be good to know the logic here.