SparseGS

SparseGS augments 3D Gaussian Splatting with depth-based priors, tailored depth rendering, a floater-pruning heuristic, and Unseen Viewpoint Regularization, letting it overcome “floaters” and background collapse when training views are scarce. Tested on Mip-NeRF360, LLFF, and DTU, it still trains quickly and renders in real time while reconstructing unbounded or forward-facing scenes from as few as 12 and 3 input images, respectively.

Web: https://formycat.github.io/SparseGS-Real-Time-360-Sparse-View-Synthesis-using-Gaussian-Splatting/
Paper: SparseGS: Real-Time 360° Sparse View Synthesis using Gaussian Splatting
Authors: Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, Achuta Kadambi

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 19.26 0.553 0.367 32m 50s 10.64 GB
bicycle n12 17.10 0.342 0.495 31m 41s 10.59 GB
flowers n12 14.08 0.241 0.582 32m 51s 10.45 GB
treehill n12 15.46 0.344 0.577 32m 57s 10.96 GB
stump n12 18.15 0.343 0.542 28m 11s 10.01 GB
kitchen n12 20.80 0.739 0.338 47m 38s 11.64 GB
bonsai n12 18.23 0.685 0.443 42m 58s 11.42 GB
counter n12 18.05 0.643 0.465 46m 29s 11.35 GB
room n12 20.40 0.733 0.443 45m 35s 11.41 GB
garden n24 23.74 0.744 0.243 41m 2s 12.36 GB
bicycle n24 19.80 0.479 0.424 39m 57s 12.54 GB
flowers n24 16.39 0.344 0.530 39m 35s 11.52 GB
treehill n24 18.81 0.444 0.529 38m 29s 11.79 GB
stump n24 20.03 0.443 0.481 36m 19s 12.22 GB
kitchen n24 24.20 0.839 0.258 57m 15s 12.31 GB
bonsai n24 22.76 0.818 0.381 52m 36s 12.52 GB
counter n24 22.64 0.778 0.382 55m 36s 12.13 GB
room n24 23.71 0.800 0.400 55m 29s 12.23 GB
Average 19.65 0.573 0.438 42m 5s 11.56 GB