Spaces:

---

usm3d

/

S23DR2026

like 10

Running on CPU Upgrade

Hi ! while training on usm3d/hoho22k_2026_trainval, I noticed that for some validation buildings the wf_vertices don't line up with the images when projected using the cameras that come with the sample.

To project, I used the cameras inside the colmap zip (via pycolmap.Reconstruction(...).images[i].cam_from_world()), since those are the cameras available for every view (the top-level K/R/t are zero when pose_only_in_colmap == True).

Example: building index 77 in the validation split. Red dots are the projected wf_vertices, green dots are the projected colmap.points3D:

The green points cover the building nicely (as expected since they came from these images), but the red vertices sit visibly below / off the actual roof edges — not something that occlusion alone explains.

Doing the same for every validation building and measuring the 3D distance from each wf_vertex to the nearest colmap.points3D.xyz:

For 6 / 170 val buildings the median GT→COLMAP distance is > 1 m:

Per-view projection overlays for all 6 are here, with 4 normal ones for comparison:

This one is less than above things, but it's shifted.

My guess is wf_vertices might be in the BPO frame while colmap.points3D is in the SfM frame, and for these 6 buildings the two frames aren't close.

A couple of questions:

1. Is this a known issue, or am I doing something wrong in how I'm getting the cameras?
2. If wf_vertices really is in a different frame, is there a per-sample transform we're supposed to apply that I missed?
3. Is the same mismatch present in the train, test(for leaderboard) split?

Thanks!