how can i mitgate this issue when scanning area ( walls not overalping )

[MohamedKHALILRouissi]

i did take 2 scans and tried the inference the results where 2 unparallel walls what can be the issue here

[Nik-V9]

Hi, is this the CC-BY-NC model or Apache one?

[MohamedKHALILRouissi]

hello thanks for the reply i used this facebook/map-anything

[MohamedKHALILRouissi]

What I’m currently trying to do

I take regular photos using a tripod, rotating the camera every 20 degrees to ensure good overlap, then stitch them into a 360° panorama using Hugin.

I split the stitched panorama into multiple perspective views with a 100° field of view.

I pass all these views to the model for inference.

Some of the results are good, but I consistently run into issues where:

Parts of the floor are duplicated

Walls are not properly connected

Is there something I can do to improve this?
Am I missing an important preprocessing or alignment step that should be applied before running the model? ( close to matterport features for phone scanning )

[cr333]

@MohamedKHALILRouissi:
If you want to reconstruct a 360° image, you might also want to look at dedicated methods such as Depth Any Camera (CVPR 2025) or Depth Any Panoramas (arXiv 2025).

[MohamedKHALILRouissi]

@cr333 even with the flow that i have now the reconstruction considered good but my issue is some time am having walls that are not connected ( so basically still doing mono depth )

[cr333]

@MohamedKHALILRouissi, I understand your issue, but MapAnything is not trained for monocular depth estimation, so dedicated approaches tend to work better for this specific task.

One way to align MapAnything's per-view reconstructions, and thus make them more consistent on the walls, would be using 360MonoDepth.