Supplementary material of DVGOv2
Results breakdown and rendered videos by DVGOv2.
Go to: project page.
Go to: DVGOv2 technical report.
mip-NeRF-360 dataset
\begin{table*}[htpb]
\centering
\begin{tabular}{l|c|ccccccc}
\hline
Method & Avg. & {\it bicycle} & {\it bonsai} & {\it counter} & {\it garden} & {\it kitchen} & {\it room} & {\it stump} \\
\hline\hline
\multicolumn{9}{@{}l}{\rule{0pt}{3ex}\bf Training minutes$\downarrow ^{\text{(*)}}$} \\
\hline
NeRF & 249.6 & - & - & - & - & - & - & - \\
NeRF++ & 567.0 & - & - & - & - & - & - & - \\
mip-NeRF 360 & 358.2 & - & - & - & - & - & - & - \\
DVGOv2 & 15.6 & 16.5 & 15.0 & 15.8 & 15.1 & 14.3 & 15.3 & 17.2 \\
\hline
\multicolumn{9}{@{}l}{\rule{0pt}{3ex}\bf PSNR$\uparrow$} \\
\hline
NeRF & 24.85 & 21.76 & 26.81 & 25.67 & 23.11 & 26.31 & 28.56 & 21.73 \\
NeRF++ & 26.21 & 22.64 & 29.15 & 26.38 & 24.32 & 27.80 & 28.87 & 24.34 \\
mip-NeRF 360 & 28.94 & 23.99 & 33.06 & 29.51 & 26.10 & 32.13 & 31.53 & 26.27 \\
DVGOv2 & 25.42 & 22.12 & 27.80 & 25.76 & 24.34 & 26.00 & 28.33 & 23.59 \\
\hline
\multicolumn{9}{@{}l}{\rule{0pt}{3ex}\bf SSIM$\uparrow$} \\
\hline
NeRF & 0.659 & 0.455 & 0.792 & 0.775 & 0.546 & 0.749 & 0.843 & 0.453 \\
NeRF++ & 0.729 & 0.526 & 0.876 & 0.802 & 0.635 & 0.816 & 0.852 & 0.594 \\
mip-NeRF 360 & 0.837 & 0.666 & 0.937 & 0.894 & 0.785 & 0.920 & 0.914 & 0.741 \\
DVGOv2 & 0.695 & 0.481 & 0.829 & 0.783 & 0.628 & 0.710 & 0.852 & 0.580 \\
\hline
\multicolumn{9}{@{}l}{\rule{0pt}{3ex}\bf LPIPS$\downarrow$ {\footnotesize (Vgg)}} \\
\hline
NeRF & 0.426 & 0.536 & 0.398 & 0.394 & 0.415 & 0.335 & 0.353 & 0.551 \\
NeRF++ & 0.348 & 0.455 & 0.291 & 0.351 & 0.331 & 0.260 & 0.335 & 0.416 \\
mip-NeRF 360 & 0.208 & 0.298 & 0.184 & 0.204 & 0.175 & 0.126 & 0.210 & 0.259 \\
DVGOv2 & 0.429 & 0.510 & 0.401 & 0.434 & 0.376 & 0.400 & 0.425 & 0.460 \\
\hline
\multicolumn{9}{l}{\footnotesize$^{\text{(*)}}$~NeRF, NeRF++, and mip-NeRF 360 are measured on TPU v2 32 cores. DVGOv2 is measured on a RTX 2080Ti GPU.}
\end{tabular}
\end{table*}
LLFF dataset
\begin{table*}[htpb]
\centering
\begin{tabular}{l|c|cccccccc}
\hline
Method & Avg. & {\it fern} & {\it flower} & {\it fortress} & {\it horns} & {\it leaves} & {\it orchids} & {\it room} & {\it trex} \\
\hline\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf Training minutes$\downarrow ^{\text{(*)}}$} \\
\hline
NeRF & hours & - & - & - & - & - & - & - & - \\
Plenoxels & 24.2 & 23.7 & 22.0 & 31.2 & 26.3 & 13.3 & 23.4 & 28.8 & 24.8 \\
TensoRF & 25.7 & - & - & - & - & - & - & - & - \\
DVGOv2 & 10.9 & 11.4 & 10.8 & 10.5 & 10.9 & 9.9 & 12.2 & 11.0 & 10.3 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf PSNR$\uparrow$} \\
\hline
NeRF & 26.50 & 25.17 & 27.40 & 31.16 & 27.45 & 20.92 & 20.36 & 32.70 & 26.80 \\
Plenoxels & 26.29 & 25.46 & 27.83 & 31.09 & 27.58 & 21.41 & 20.24 & 30.22 & 26.48 \\
TensoRF & 26.73 & 25.27 & 28.60 & 31.36 & 28.14 & 21.30 & 19.87 & 32.35 & 26.97 \\
DVGOv2 & 26.34 & 25.08 & 27.62 & 30.44 & 27.59 & 21.00 & 20.33 & 31.53 & 27.17 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf SSIM$\uparrow$} \\
\hline
NeRF & 0.811 & 0.792 & 0.827 & 0.881 & 0.828 & 0.690 & 0.641 & 0.948 & 0.880 \\
Plenoxels & 0.839 & 0.832 & 0.862 & 0.885 & 0.857 & 0.760 & 0.687 & 0.937 & 0.890 \\
TensoRF & 0.839 & 0.814 & 0.871 & 0.897 & 0.877 & 0.752 & 0.649 & 0.952 & 0.900 \\
DVGOv2 & 0.838 & 0.825 & 0.854 & 0.885 & 0.867 & 0.745 & 0.679 & 0.946 & 0.904 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf LPIPS$\downarrow$ {\footnotesize (Vgg)}} \\
\hline
NeRF & 0.250 & 0.280 & 0.219 & 0.171 & 0.268 & 0.316 & 0.321 & 0.178 & 0.249 \\
Plenoxels & 0.210 & 0.224 & 0.179 & 0.180 & 0.231 & 0.198 & 0.242 & 0.192 & 0.238 \\
TensoRF & 0.204 & 0.237 & 0.169 & 0.148 & 0.196 & 0.217 & 0.278 & 0.167 & 0.221 \\
DVGOv2 & 0.197 & 0.210 & 0.179 & 0.159 & 0.207 & 0.202 & 0.238 & 0.175 & 0.209 \\
\hline
\multicolumn{10}{l}{\footnotesize$^{\text{(*)}}$~Plenoxels uses a Titan RTX; TensoRF uses a Telsa V100; DVGOv2 uses the lowest spec RTX 2080Ti.}
\end{tabular}
\end{table*}
Tanks&Temples dataset
\begin{table*}[htpb]
\centering
\begin{tabular}{l|c|cccc}
\hline
Method & Avg. & {\it m60} & {\it playground} & {\it train} & {\it truck} \\
\hline\hline
\multicolumn{6}{@{}l}{\rule{0pt}{3ex}\bf Training minutes$\downarrow ^{\text{(*)}}$} \\
\hline
NeRF++ & hours & - & - & - & - \\
Plenoxels & 27.3 & 25.5 & 26.3 & 29.5 & 28.0 \\
DVGOv2 & 16.0 & 14.1 & 16.6 & 18.0 & 15.4 \\
\hline
\multicolumn{6}{@{}l}{\rule{0pt}{3ex}\bf PSNR$\uparrow$} \\
\hline
NeRF++ & 20.49 & 18.49 & 22.93 & 17.77 & 22.77 \\
Plenoxels & 20.40 & 17.93 & 23.03 & 17.97 & 22.67 \\
DVGOv2 & 20.10 & 17.53 & 22.70 & 17.92 & 22.24 \\
\hline
\multicolumn{6}{@{}l}{\rule{0pt}{3ex}\bf SSIM$\uparrow$} \\
\hline
NeRF++ & 0.648 & 0.650 & 0.672 & 0.558 & 0.712 \\
Plenoxels & 0.696 & 0.687 & 0.712 & 0.629 & 0.758 \\
DVGOv2 & 0.653 & 0.652 & 0.675 & 0.570 & 0.716 \\
\hline
\multicolumn{6}{@{}l}{\rule{0pt}{3ex}\bf LPIPS$\downarrow$ {\footnotesize (Vgg)}} \\
\hline
NeRF++ & 0.478 & 0.481 & 0.477 & 0.531 & 0.424 \\
Plenoxels & 0.420 & 0.439 & 0.435 & 0.443 & 0.364 \\
DVGOv2 & 0.477 & 0.483 & 0.492 & 0.514 & 0.418 \\
\hline
\multicolumn{6}{l}{\footnotesize$^{\text{(*)}}$~Plenoxels uses a Titan RTX, while DVGOv2 uses the lower spec RTX 2080Ti.}
\end{tabular}
\end{table*}
Synthetic-NeRF dataset
\begin{table*}[htpb]
\centering
\begin{tabular}{l|c|cccccccc}
\hline
Method & Avg. & {\it chair} & {\it drums} & {\it ficus} & {\it hotdog} & {\it lego} & {\it materials} & {\it mic} & {\it ship} \\
\hline\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf Training minutes$\downarrow ^{\text{(*)}}$} \\
\hline
DVGO & 14.2 & 12.5 & 12.0 & 13.8 & 15.5 & 13.2 & 15.4 & 11.0 & 20.2 \\
Plenoxels & 11.1 & 9.6 & 9.8 & 8.8 & 12.5 & 10.8 & 11.0 & 8.2 & 18.0 \\
TensoRF & 17.6 & - & - & - & - & - & - & - & - \\
Instant-NGP & 5 & - & - & - & - & - & - & - & - \\
DVGOv2 & 6.8 & 5.9 & 5.7 & 6.2 & 7.3 & 6.7 & 8.8 & 4.6 & 9.2 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf PSNR$\uparrow$} \\
\hline
DVGO & 31.95 & 34.09 & 25.44 & 32.78 & 36.74 & 34.64 & 29.57 & 33.20 & 29.13 \\
Plenoxels & 31.71 & 33.98 & 25.35 & 31.83 & 36.43 & 34.10 & 29.14 & 33.26 & 29.62 \\
TensoRF & 33.14 & 35.76 & 26.01 & 33.99 & 37.41 & 36.46 & 30.12 & 34.61 & 30.77 \\
Instant-NGP & 33.18 & 35.00 & 26.02 & 33.51 & 37.40 & 36.39 & 29.78 & 36.22 & 31.10 \\
DVGOv2 & 32.76 & 35.33 & 25.83 & 33.57 & 37.24 & 35.73 & 29.68 & 34.44 & 30.28 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf SSIM$\uparrow$} \\
\hline
DVGO & 0.957 & 0.977 & 0.930 & 0.978 & 0.980 & 0.976 & 0.951 & 0.983 & 0.879 \\
Plenoxels & 0.958 & 0.977 & 0.933 & 0.976 & 0.980 & 0.976 & 0.949 & 0.985 & 0.890 \\
TensoRF & 0.963 & 0.985 & 0.937 & 0.982 & 0.982 & 0.983 & 0.952 & 0.988 & 0.895 \\
Instant-NGP & - & - & - & - & - & - & - & - & - \\
DVGOv2 & 0.962 & 0.983 & 0.936 & 0.982 & 0.982 & 0.980 & 0.951 & 0.988 & 0.892 \\
\hline
\multicolumn{10}{@{}l}{\rule{0pt}{3ex}\bf LPIPS$\downarrow$ {\footnotesize (Vgg)}} \\
\hline
DVGO & 0.053 & 0.027 & 0.077 & 0.024 & 0.034 & 0.028 & 0.058 & 0.017 & 0.161 \\
Plenoxels & 0.049 & 0.031 & 0.067 & 0.026 & 0.037 & 0.028 & 0.057 & 0.015 & 0.134 \\
TensoRF & 0.047 & 0.022 & 0.073 & 0.022 & 0.032 & 0.018 & 0.058 & 0.015 & 0.138 \\
Instant-NGP & - & - & - & - & - & - & - & - & - \\
DVGOv2 & 0.046 & 0.020 & 0.068 & 0.020 & 0.029 & 0.021 & 0.054 & 0.012 & 0.141 \\
\hline
\multicolumn{10}{l}{\footnotesize \makecell[l]{$^{\text{(*)}}$~Plenoxels uses a Titan RTX; TensoRF uses a Telsa V100; Instant-NGP uses a RTX 3090;\\~~~~~DVGO and DVGOv2 uses the lowest spec RTX 2080Ti.}}
\end{tabular}
\end{table*}
Tanks&Temple (bounded ver.) dataset
\begin{table*}[htpb]
\centering
\begin{tabular}{l|c|ccccc}
\hline
Method & Avg. & {\it Ignatius} & {\it Truck} & {\it Barn} & {\it Caterpillar} & {\it Family} \\
\hline\hline
\multicolumn{7}{@{}l}{\rule{0pt}{3ex}\bf Training minutes$\downarrow ^{\text{(*)}}$} \\
\hline
DVGO & 17.7 & 15.7 & 17.8 & 22.0 & 17.7 & 15.3 \\
Plenoxels & - & - & - & - & - & - \\
TensoRF & - & - & - & - & - & - \\
DVGOv2 & 9.5 & 8.1 & 9.2 & 12.6 & 9.8 & 8.0 \\
\hline
\multicolumn{7}{@{}l}{\rule{0pt}{3ex}\bf PSNR$\uparrow$} \\
\hline
DVGO & 28.41 & 28.16 & 27.15 & 27.01 & 26.00 & 33.75 \\
Plenoxels & 27.43 & 27.51 & 26.59 & 26.07 & 24.64 & 32.33 \\
TensoRF & 28.56 & 28.34 & 27.14 & 27.22 & 26.19 & 33.92 \\
DVGOv2 & 28.69 & 28.37 & 27.46 & 27.30 & 26.11 & 34.21 \\
\hline
\multicolumn{7}{@{}l}{\rule{0pt}{3ex}\bf SSIM$\uparrow$} \\
\hline
DVGO & 0.911 & 0.944 & 0.906 & 0.838 & 0.906 & 0.962 \\
Plenoxels & 0.906 & 0.943 & 0.901 & 0.829 & 0.902 & 0.956 \\
TensoRF & 0.920 & 0.948 & 0.914 & 0.864 & 0.912 & 0.965 \\
DVGOv2 & 0.918 & 0.948 & 0.915 & 0.850 & 0.912 & 0.967 \\
\hline
\multicolumn{7}{@{}l}{\rule{0pt}{3ex}\bf LPIPS$\downarrow$ {\footnotesize (Vgg)}} \\
\hline
DVGO & 0.155 & 0.083 & 0.160 & 0.294 & 0.167 & 0.070 \\
Plenoxels & 0.162 & 0.102 & 0.163 & 0.303 & 0.166 & 0.078 \\
TensoRF & 0.140 & 0.078 & 0.145 & 0.252 & 0.159 & 0.064 \\
DVGOv2 & 0.143 & 0.079 & 0.143 & 0.275 & 0.156 & 0.060 \\
\hline
\multicolumn{7}{l}{\footnotesize \makecell[l]{$^{\text{(*)}}$~Plenoxels uses a Titan RTX; TensoRF uses a Telsa V100;\\~~~~~DVGO and DVGOv2 uses the lowest spec RTX 2080Ti.}}
\end{tabular}
\end{table*}