Support backgrounds in Gaussian Rasterization

[harrism]

Fixes #80

This adds support for backgrounds (per-camera background color / feature) to fVDB's Gaussian Splatting. The kernels already supported backgrounds, but it was not plumbed into the public interface. This PR:

1. Adds gtests to verify kernel correctness
2. Plumbs backgrounds parameters all the way up to the Python public interface.
3. Adds pytests that exercise backgrounds.

Signed-off-by: Mark Harris mharris@nvidia.com

[Copilot]

Pull Request Overview

This PR adds support for background colors/features to Gaussian Splatting rasterization. The feature allows per-camera background colors to be blended with transparent pixels during rendering, addressing issue #80.

Key changes:

• Added backgrounds parameter to all rendering functions (dense and sparse, forward and backward passes)
• Implemented comprehensive test coverage for background blending in both Python and C++ tests
• Fixed a spelling error in an existing comment

Reviewed Changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tests/unit/test_gsplat.py	Added comprehensive Python tests for background rendering in multiple scenarios
src/tests/GaussianRasterizeForwardTest.cpp	Added C++ tests for forward pass with backgrounds; fixed comment typo
src/tests/GaussianRasterizeBackwardTest.cpp	Added C++ tests for backward pass with backgrounds
src/python/GaussianSplatBinding.cpp	Added backgrounds parameter to Python bindings
src/fvdb/detail/ops/gsplat/GaussianRasterizeForward.h	Updated function signatures to include backgrounds parameter
src/fvdb/detail/ops/gsplat/GaussianRasterizeForward.cu	Plumbed backgrounds through forward implementation
src/fvdb/detail/ops/gsplat/GaussianRasterizeBackward.h	Updated backward pass signatures to include backgrounds
src/fvdb/detail/ops/gsplat/GaussianRasterizeBackward.cu	Plumbed backgrounds through backward implementation
src/fvdb/detail/autograd/GaussianRender.h	Added backgrounds to autograd function signature
src/fvdb/detail/autograd/GaussianRender.cpp	Implemented backgrounds handling in autograd
src/fvdb/GaussianSplat3d.h	Added backgrounds parameter to public API methods
src/fvdb/GaussianSplat3d.cpp	Updated rendering implementations to support backgrounds
fvdb/gaussian_splatting.py	Added backgrounds parameter to Python API
fvdb/_fvdb_cpp.pyi	Updated type stubs for backgrounds parameter
fvdb/init.pyi	Updated type stubs for gaussian_render_jagged
fvdb/init.py	Added backgrounds parameter to gaussian_render_jagged wrapper

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[fwilliams]

A few small questions but overall looks good. We may want to have a mode where we pass in a full background image rather than a fixed color which would let us implement e.g. a learnable skybox, but this can be a different PR when we decide to actually implement that feature.

[swahtz]

This is great, thanks for plumbing that in and making the tests very complete. I like that the tests also include validating the expected alpha composited background result. I had thought that 'background' was a whole image-per-camera but now it makes sense it's just for training with random background values.