Normal supervision - Is normal properly expressed in camera coordinates before rendering ?

[gaetan-landreau]

Hi ! 👋🏼
I recently skimmed through your code and this particular line made me a bit lost ..

dn-splatter/dn_splatter/dn_model.py

Line 556 in 367c36d

normals = normals @ camera.camera_to_worlds.squeeze(0)[:3, :3]

As comment suggest that it "convert normals from world space to camera space", I would have rather taken the inverse of camera.camera_to_worlds to express the normals, defined in world coordinate, into the camera one. Am I missing something ? 🥲

[sevashasla]

Hello!
If I am not mistaken the code is correct. Normals are vectors-rows but not vectors-columns. Therefore the result in vector-s-columns should be (w2c @ normal.T) = (c2w.T @ normal.T)
We want to get vectors-rows back, then
(c2w.T @ normal.T).T = normal @ c2w