tink/nerfstudio
4
0
Fork 0
mirror of https://github.com/nerfstudio-project/nerfstudio.git synced 2026-11-11 18:44:21 +08:00
Code
25a00a09ab
nerfstudio/docs/nerfology/methods/splat.md
AdamRashid96 a8d5dc9ae9
Updates to docs for gaussian splatting (#2782) 
* additions to docs

---------

Co-authored-by: Justin Kerr <justin.g.kerr@gmail.com>
2024-01-18 11:17:54 -08:00

4.2 KiB
Raw Blame History

Gaussian Splatting

Real-Time Radiance Field Rendering

:color: primary
:outline:
Paper Website

3D Gaussian Splatting was proposed in SIGGRAPH 2023 from INRIA, and is a completely different method of representing radiance fields by explicitly storing a collection of 3D volumetric gaussians. These can be "splatted", or projected, onto a 2D image provided a camera pose, and rasterized to obtain per-pixel colors. Because rasterization is very fast on GPUs, this method can render much faster than neural representations of radiance fields.

Installation

:color: primary
:outline:
GSplat

Nerfstudio uses gsplat as its gaussian rasterization backend, an in-house re-implementation which is designed to be more developer friendly. This can be installed with pip install gsplat. The associated CUDA code will be compiled the first time gaussian splatting is executed. Some users with PyTorch 2.0 have experienced issues with this, which can be resolved by either installing gsplat from source, or upgrading torch to 2.1.

Data

Gaussian Splatting works much better if you initialize it from pre-existing geometry, such as SfM points from COLMAP. COLMAP datasets or datasets from ns-process-data will automatically save these points and initialize gaussians on them. Other datasets currently do not support initialization, and will initialize gaussians randomly. Initializing from other data inputs (i.e. depth from phone app scanners) may be supported in the future.

Because gaussian splatting trains on full images instead of bundles of rays, there is a new datamanager in full_images_datamanager.py which undistorts input images, caches them, and provides single images at each train step.

Running the Method

To run gaussian splatting, run ns-train gaussian-splatting --data <data>. Just like NeRF methods, the splat can be interactively viewed in the web-viewer, loaded from a checkpoint, rendered, and exported.

Quality and Regularization

The default settings provided maintain a balance between speed, quality, and splat file size, but if you care more about quality than training speed or size, you can decrease the alpha cull threshold (threshold to delete translucent gaussians) and disable culling after 15k steps like so: ns-train gaussian-splatting --pipeline.model.cull_scale_thresh=0.005 --pipeline.model.continue_cull_post_densification=False --data <data>

A common artifact in splatting is long, spikey gaussians. PhysGaussian proposes an example of a scale-regularizer that encourages gaussians to be more evenly shaped. To enable this, set the use_scale_regularization flag to True.

Details

For more details on the method, see the original paper. Additionally, for a detailed derivation of the gradients used in the gsplat library, see here.

Exporting splats

Gaussian splats can be exported as a .ply file which are ingestable by a variety of online web viewers. You can do this via the viewer, or ns-export gaussian-splat --load-config <config> --output-dir exports/splat. Currently splats can only be exported from trained splats, not from nerfacto.

Nerfstudio gaussian splat exports currently supports multiple third-party splat viewers:

FAQ

  • Can I export a mesh or pointcloud? Currently these export options are not supported, but may become in the future. Contributions are always welcome!
  • Can I render fisheye, equirectangular, orthographic images? Currently, no. Gaussian splatting assumes a perspective camera for its rasterization pipeline. Implementing other camera models is of interest but not currently planned.