Skip to content

Latest commit

 

History

History
33 lines (27 loc) · 3.13 KB

EXPLORATION.md

File metadata and controls

33 lines (27 loc) · 3.13 KB

Exploration For Better 3D Gaussian Splatting

  • AbsGrad: Uses absolute gradients in the image plane as the criterion for pruning. See this paper for more details.
  • Antialiasing: Applies a low pass filter on the projected covariance and scale the opacity accordingly. See this paper for more details. It might slightly hurt the metrics on in-distribution views but seem to improve the visual quality on view out of training distribution.
Garden at 7k steps (TITAN RTX) T(train) T(render) Memory SSIM PSNR LPIPS #GS.
default args 7m07s 0.021s/im 7.54 GB 0.8332 26.29 0.123 4.46M
--absgrad --grow_grad2d 8e-4 5m50s 0.012s/im 3.80 GB 0.8365 26.44 0.121 2.17M
--absgrad --grow_grad2d 8e-4 (30k) -- 0.013s/im 4.04 GB 0.8639 27.33 0.079 2.35M
--antialiased 6m43s 0.020s/im 6.74 GB 0.8265 26.13 0.137 3.99M
U1 at 7k steps (RTX 2080 Ti) T(train) T(render) Memory SSIM PSNR LPIPS #GS.
default args 7m39s 0.013s/im 4.94 GB 0.6102 20.69 0.615 2.47M
default args (30k) -- 0.019s/im -- 0.7518 24.67 0.385 4.18M
--absgrad --grow_grad2d 8e-4 7m16s 0.011s/im 3.41 GB 0.6055 20.29 0.636 1.72M
--absgrad --grow_grad2d 8e-4 (30k) -- 0.014s/im 4.15 GB 0.7494 24.65 0.390 2.37M
--absgrad --grow_grad2d 6e-4 8m58s 0.011s/im 4.42 GB 0.5966 19.58 0.654 2.21M
--absgrad --grow_grad2d 6e-4 (30k) -- 0.016s/im 5.09 GB 0.7439 24.28 0.400 2.92M
U4 at 7k steps (RTX 2080 Ti) T(train) T(render) Memory SSIM PSNR LPIPS #GS.
--grow_grad2d 5e-5 7m30s 0.014s/im 1.68 GB 0.6271 20.86 0.583 0.61M
--grow_grad2d 5e-5 (30k) -- 0.026s/im 4.21 GB 0.7402 24.05 0.299 2.44M
--absgrad --grow_grad2d 2e-4 8m30s 0.018s/im 2.21 GB 0.6251 20.68 0.587 0.89M
--absgrad --grow_grad2d 2e-4 (30k) -- 0.030s/im 5.25 GB 0.7442 24.12 0.291 2.62M

Note: default args means running CUDA_VISIBLE_DEVICES=0 python simple_trainer.py default --data_dir <DATA_DIR> with:

  • Garden (Source): --result_dir results/garden
  • U1 (a.k.a University 1 from Source): --result_dir results/u1 --data_factor 1 --grow_scale3d 0.001
  • U4 (a.k.a University 4 from Source): --result_dir results/u4 --data_factor 1 --grow_scale3d 0.01 --refine_every 500