[DRAFT] Automatic circular camera path rendering #3314
Conversation
nerfstudio/viewer/render_panel.py
Outdated
| fars=torch.tensor(100).view(1, 1), | ||
| ).to("cuda") | ||
|
|
||
| # Get the distance/depth to the intersection --> calculate 3D position of the click |
There was a problem hiding this comment.
This is a good start! but this part should use the more general get_outputs_for_camera() function (a lot of the sampling, scaling stuff is method-specific to nerfacto). As long as the method outputs a 'depth' value it should work with this ray deprojection approach
There was a problem hiding this comment.
This way, you would render a depth image from the viewer camera and deproject the click point with the intrinsics matrix + rendered depth, meaning it doesn't matter what the rendering backend is.
There was a problem hiding this comment.
One thing to note is that the 'depth' in splatfacto is actually z-depth and not ray-depth, so the math would need to be different for the two methods
There was a problem hiding this comment.
Ok that makes sense! Thanks for the feedback appreciate it!!
Interface to generate a circular camera path! Object avoidance is WIP, but a vanilla circular path for now can be useful for object comparisons.
Screen.Recording.2024-07-17.at.4.55.47.PM.1.1.mov
TODO: