PCAsplat Cleans Up Gaussian Splat "Floaters" With Local PCA on Three Benchmarks

Gaussian Splatting has long outgrown pretty rendering, but geometry remains its weak spot. The team behind PCAsplat proposes a regularizer based on differentiable local principal component analysis (PCA) that directly moves Gaussian centers toward the real surface of the scene. The method was tested on three datasets: DTU, Tanks and Temples, and NeRF Synthetic. According to the authors, the splats approximate the ground-truth surface more accurately, and the number of floating artifacts drops significantly. The authors say they will release the code.

Why Gaussians drift away

Standard 3D Gaussian Splatting training relies on rasterization losses. Gradients reach only those Gaussians that fall within the rays of the selected cameras and contribute to a pixel. If a splat is occluded by another object or barely visible from the current viewpoint, the signal about its position is weak or zero. Such a splat stays wherever it ended up in the early iterations and eventually hangs in the air near the surface. These are floaters: floating artifacts that corrupt the point cloud, the mesh, and any downstream operation on the geometry.

The problem is systemic: photometric error does not know where the surface is. It only checks whether the final frame looks like the photograph.

How PCAsplat works

The authors add losses that do not depend on the camera. For each Gaussian, the method takes a neighborhood of the nearest centers and computes a local PCA. The operation is differentiable, so the gradient flows back into the center positions, including those of splats that do not participate in the current view.

The regularizer has two parts:

• Eigenvalue regularization. It pushes Gaussians toward the surface and enforces isotropic coverage of the tangent plane: in effect, a small eigenvalue along the normal and two comparable ones in the other directions.
• Normal alignment. The normal of each Gaussian is made consistent with the normal estimated by PCA from its neighborhood, which yields a uniform orientation.

The method does not replace the rasterization losses. It works alongside them as an additional constraint.

What the experiments showed

The abstract contains no numerical tables, so the size of the improvement cannot be judged yet. The following claims are qualitative:

• Benchmarks: DTU, Tanks and Temples, NeRF Synthetic.
• The splats better approximate a point sample from the ground-truth surface.
• There are significantly fewer floating artifacts.
• Novel view synthesis and mesh extraction remain at a competitive level, meaning the authors do not promise a noticeable leap on classic rendering metrics.
• Surface-aligned splats are suitable for point cloud segmentation and direct Poisson reconstruction, without intermediate workaround steps.

The last point is the most practical. Poisson reconstruction requires clean points with consistent normals, and that is exactly what the regularizer tries to deliver at the end of training.

Who this is useful for

If you build a scanning, photogrammetry, or digital twin pipeline on Gaussian splats, the main value of the method is that the cloud of centers can be used as geometry, not just as a rendering asset. For robotics and AR, this means less manual cleanup before building meshes and collision geometry.

Before adopting it, wait for the code and check three things. First, how much time neighbor search and PCA add per iteration on scenes with millions of Gaussians. Second, how the method behaves on scenes with thin structures and specular highlights, where the local-plane assumption breaks down. Third, how well the geometry gains hold up on your own data, not just on three academic datasets. For now, it is a neat and logical idea for those who care about geometry as much as the picture.