search

EPFL-VILAB / omnidata

A Scalable Pipeline for Making Steerable Multi-Task Mid-Level Vision Datasets from 3D Scans [ICCV 2021]
Other
395 stars 49 forks source link

Pose consistency in image sequences #46

Closed VitorGuizilini-TRI closed 1 year ago

VitorGuizilini-TRI commented 1 year ago

Hi, I am trying to reconstruct sequences using a global frame of reference, but it seems like there is a lot of misalignment between certain samples. Can you explain a little more how the "point" and "view" system works?

My questions is whether I can assume that all images from the same sequence (i.e., folder) follow a global frame of reference in the stored poses, so I could in principle recover the entire scene by reconstructing from all poses at the same time. It seems like that is the case for some images, but not for others (sometimes I get perfect alignment, and sometimes they don't align at all).

Thank you!

alexsax commented 1 year ago

For each building, we sample points on the mesh. Then we check which cameras have line-of-sight to the point. For each combination of building, point, visible camera, we call this a (building, point, view) triplet.

The R, T, K matrices returned by the dataloader should be in a building-wide global coordinate frame. For which buildings, points, and views are you seeing a misalignment? Are you seeing this issue with the dataloader provided here?

VitorGuizilini-TRI commented 1 year ago

Never mind, I found the issue here, it's all good now! Thanks!