How are camera translations scaled?

[NagabhushanSN95]

Hi,

I've a dataset of images with given camera intrinsics and extrinsics (pose). So, I'm trying to generate the transforms_train.json file without needing to run colmap and other stuff. To do this, I'm trying to figure out how transforms_train.json is created based on colmap sparse reconstruction for the ScanNet scenes. I figured out the relation between rotation matrices, however, I'm not able to figure out how translation is scaled. I found different scaling factors for the 5 scenes. I tried to understand the C++ code that generates the transforms_train.json file, but I'm not used to C++ and hence couldn't figure it out.

Can you please tell me how to compute the scaling factor for the camera translation?

PS: I also noticed that, unlike the original NeRF which scales translation so that nearest depth becomes unity, you do not do such a scaling here.

[barbararoessle]

Here is an explanation on the content of transforms_train.json and the definition of the transformation matrix. Does this answer your question?

In the preprocessing on the ScanNet SfM reconstructions we apply a scaling to the reconstruction (sparse depth and poses), because SfM reconstructions are only upto a scaling factor. This scaling allows to compute depth metrics against the ScanNet sensor depth.
Scaling of the sparse depth and poses is needed only when:

• you have ground truth depth to evaluate against with different scaling than the camera poses and sparse depth, or
• the depth prior network was trained on a different range of sparse depth.

[NagabhushanSN95]

Thanks, Barbara! This helps :)