Gaussian splatting support for Aria #2785
Conversation
| undist_h = 500 | ||
| undist_w = 500 | ||
| undist_K = torch.eye(3) | ||
| undist_K[0, 0] = 150.0 # fx | ||
| undist_K[1, 1] = 150.0 # fy |
There was a problem hiding this comment.
These seems pretty small from a quick eye glance! is this just hardcoded here for testing/prototyping?
There was a problem hiding this comment.
Should not come from the camera calibration itself so it would work for various Aria recording data profile
There was a problem hiding this comment.
agree, isn't it weird for these to be hardcoded?
There was a problem hiding this comment.
Yeah this was test code; I added some heuristics for computing this automatically! Not sure if there's a better way...
brentyi
force-pushed
the
brent/aria_gsplat
branch
from
88f9f04 to
d1783df
Compare
|
Thanks for the feedback everyone! I added logic for computing the undistorted camera parameters, and fixed some issues with the masks. Some more results: The results look cleaner now: With initialization points (filtered from MPS) overlayed: |
|
Thank you for fixing the masks in GS! d1783df LGTM! |
* Gaussian splatting support for Aria * Respect masks in splatting loss function
Added support for the Fisheye624 model in the full image dataloader. @kerrj can you check if my changes make sense?
ns-process-data arianow also saves the Aria MPS points, which seems to improve results slightly.Without points:
With points:
(this is the same capture as in #2617)
FYI @sweeneychris