How should the scale range be determined for a new data set?

dcharatan / pixelsplat

[CVPR 2024 Oral, Best Paper Runner-Up] Code for "pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction" by David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann

http://davidcharatan.com/pixelsplat/

MIT License

830 stars 56 forks source link

How should the scale range be determined for a new data set? #54

Closed caiyongqi closed 6 months ago

caiyongqi commented 6 months ago

https://github.com/dcharatan/pixelsplat/blob/f4ae9c1396ce62097b407e819321f5819d7788a9/src/model/encoder/common/gaussian_adapter.py#L65

dcharatan commented 6 months ago

The scales are expressed in terms of pixel widths in the input image. See this comment for more details.

For a new dataset, I would start with the current scale range, then update it if the Gaussians seem too large or small on average. Defining the scale in terms of pixel widths is supposed to make the scales invariant to image resolution, but I suspect that larger images will require larger max scales, since it can be helpful to have large Gaussians even for high-resolution images.