DropGaussian

DropGaussian extends 3DGS with a simple regularization technique for sparse view novel view synthesis.

Web: https://github.com/DCVL-3D/DropGaussian_release
Paper: DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting
Authors: Hyunwoo Park, Gun Ryu, Wonjun Kim

Mip-NeRF 360 Sparse

Modified Mip-NeRF 360 dataset with small train set (12 or 24) views. The dataset is used to evaluate sparse-view NVS methods.

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
garden n12 20.74 0.579 0.421 4m 36s 4.21 GB
bicycle n12 19.46 0.425 0.523 4m 21s 3.97 GB
flowers n12 16.39 0.296 0.615 4m 25s 3.96 GB
treehill n12 17.39 0.413 0.610 4m 23s 4.21 GB
stump n12 19.84 0.406 0.571 4m 5s 3.88 GB
kitchen n12 22.83 0.792 0.306 6m 60s 5.49 GB
bonsai n12 21.23 0.786 0.404 6m 14s 5.60 GB
counter n12 20.07 0.725 0.425 6m 25s 5.53 GB
room n12 22.78 0.797 0.422 6m 31s 5.49 GB
garden n24 24.26 0.694 0.369 4m 35s 4.46 GB
bicycle n24 21.27 0.496 0.496 4m 29s 4.42 GB
flowers n24 17.99 0.355 0.588 4m 35s 4.21 GB
treehill n24 20.31 0.487 0.576 4m 22s 4.33 GB
stump n24 21.09 0.466 0.556 4m 16s 4.21 GB
kitchen n24 25.85 0.863 0.248 6m 49s 5.74 GB
bonsai n24 25.12 0.863 0.351 6m 15s 5.78 GB
counter n24 24.27 0.817 0.360 6m 22s 5.82 GB
room n24 25.25 0.832 0.399 6m 37s 5.74 GB
Average 21.45 0.616 0.458 5m 21s 4.84 GB