Johan Edstedt

> I could verify the observation of @Parskatt , where changing THREADS_PER_BLOCK (and kMaxPatchMatchWindowRadius) from 32 to 96 can accelerate it by ~2.8 times on one A100 GPU (although I...

Yeah it breaks it, but I dont get why. I started looking at some other stuff in the meantime :D

@ducha-aiki made some start at adding dedode, let me know issues. I'll fix the obvious linter errors and so on.

https://twitter.com/Parskatt/status/1757423036771254499

This inconsistency makes it confusing to use. To me, projection is P: 3x4, intrinsics are K: 3x3, and pose is R: 3x3 ,t: 3x1 alt. T: 4x4 ```unproject_points``` should take...

While I'm at it, why is this:  Composed on the left? I expected it would be on the right.

Basically only need to do this: https://github.com/colmap/colmap/blob/main/src/colmap/geometry/essential_matrix.cc#L97

@edgarriba The hyperparams don't seem optimized to me, do you have a reference implementation you are referring to? For example, using these gradients: https://github.com/kornia/kornia/blob/251123a1ba29bdf2e32bfcf12176d8fa68ac9454/kornia/tracking/optical_flow_lk.py#L35 and having a window size of...

Ive TAd on this for a long time, so actually a bit tired of it. Might do it, but not prio.

Nope, I think scannet++ is the way to go if you want accuracy.