Tanks and Temples
Tanks and Temples is a benchmark for image-based 3D reconstruction. The benchmark sequences were acquired outside the lab, in realistic conditions. Ground-truth data was captured using an industrial laser scanner. The benchmark includes both outdoor scenes and indoor environments. The dataset is split into three subsets: training, intermediate, and advanced.
Authors: Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, Vladlen Koltun
Instant NGP | 21.62 | 0.712 | 0.340 | 4m 27s | 4.13 GB | |
NerfStudio | 22.04 | 0.743 | 0.270 | 19m 27s | 3.74 GB | |
Gaussian Opacity Fields | 22.39 | 0.825 | 0.172 | N/A | N/A | |
Gaussian Splatting | 23.83 | 0.831 | 0.165 | 13m 48s | 6.95 GB | |
Mip-Splatting | 23.93 | 0.833 | 0.166 | 15m 56s | 7.27 GB | |
Zip-NeRF | 24.63 | 0.840 | 0.131 | 5h 44m 9s | 26.61 GB |
PSNR
Peak Signal to Noise Ratio. The higher the better.
Instant NGP | 20.67 | 21.62 | 19.44 | 15.19 | 19.09 | 17.84 | 22.59 | 24.38 | 21.82 | 21.65 | 25.82 | 26.81 | 23.33 | 20.01 | 25.90 | 21.72 | 19.92 | 20.80 | 19.40 | 23.24 | 22.85 |
NerfStudio | 20.77 | 22.68 | 20.24 | 17.84 | 17.68 | 17.06 | 24.32 | 24.60 | 24.31 | 20.85 | 26.54 | 27.57 | 24.69 | 20.43 | 26.40 | 21.71 | 20.06 | 18.11 | 20.44 | 23.21 | 23.37 |
Gaussian Opacity Fields | 23.20 | 22.84 | 21.15 | 19.92 | 16.46 | 20.29 | 22.31 | 24.76 | 23.73 | 21.80 | 28.04 | 28.47 | 23.89 | 19.69 | 25.72 | 21.78 | 19.65 | 19.60 | 20.34 | 24.31 | 22.33 |
Gaussian Splatting | 24.13 | 24.07 | 23.12 | 20.92 | 19.63 | 20.85 | 24.43 | 27.22 | 23.82 | 22.11 | 27.84 | 28.32 | 25.37 | 21.67 | 27.51 | 23.38 | 22.79 | 22.22 | 21.53 | 25.19 | 24.25 |
Mip-Splatting | 24.41 | 24.15 | 23.00 | 20.88 | 19.63 | 20.55 | 24.55 | 27.61 | 23.94 | 22.25 | 27.98 | 28.27 | 25.87 | 21.82 | 27.75 | 23.42 | 22.76 | 22.15 | 21.73 | 25.46 | 24.36 |
Zip-NeRF | 24.52 | 25.45 | 22.17 | 19.34 | 19.11 | 20.58 | 27.10 | 29.10 | 26.82 | 23.07 | 29.01 | 28.76 | 27.13 | 22.19 | 29.26 | 23.94 | 23.14 | 22.88 | 22.61 | 25.93 | 25.09 |
SSIM
Structural Similarity Index. The higher the better. The implementation matches JAX's SSIM and torchmetrics's SSIM (with default parameters).
Instant NGP | 0.761 | 0.652 | 0.640 | 0.471 | 0.668 | 0.689 | 0.761 | 0.824 | 0.784 | 0.765 | 0.832 | 0.844 | 0.696 | 0.658 | 0.772 | 0.633 | 0.650 | 0.681 | 0.613 | 0.783 | 0.770 |
NerfStudio | 0.771 | 0.705 | 0.673 | 0.648 | 0.640 | 0.678 | 0.822 | 0.851 | 0.847 | 0.768 | 0.843 | 0.858 | 0.755 | 0.693 | 0.794 | 0.666 | 0.671 | 0.632 | 0.689 | 0.793 | 0.797 |
Gaussian Opacity Fields | 0.871 | 0.818 | 0.781 | 0.761 | 0.683 | 0.794 | 0.875 | 0.901 | 0.881 | 0.833 | 0.906 | 0.910 | 0.869 | 0.796 | 0.866 | 0.791 | 0.775 | 0.726 | 0.769 | 0.862 | 0.860 |
Gaussian Splatting | 0.871 | 0.824 | 0.790 | 0.764 | 0.736 | 0.806 | 0.865 | 0.897 | 0.875 | 0.843 | 0.901 | 0.908 | 0.848 | 0.791 | 0.852 | 0.791 | 0.811 | 0.779 | 0.776 | 0.866 | 0.853 |
Mip-Splatting | 0.872 | 0.826 | 0.791 | 0.768 | 0.731 | 0.805 | 0.872 | 0.899 | 0.879 | 0.844 | 0.904 | 0.908 | 0.861 | 0.795 | 0.855 | 0.790 | 0.812 | 0.779 | 0.780 | 0.870 | 0.857 |
Zip-NeRF | 0.877 | 0.835 | 0.790 | 0.746 | 0.718 | 0.805 | 0.889 | 0.915 | 0.897 | 0.849 | 0.912 | 0.909 | 0.880 | 0.814 | 0.884 | 0.802 | 0.807 | 0.780 | 0.789 | 0.875 | 0.864 |
LPIPS
Learned Perceptual Image Patch Similarity. The lower the better. The implementation uses AlexNet backbone and matches lpips pip package with checkpoint version 0.1
Instant NGP | 0.429 | 0.352 | 0.448 | 0.606 | 0.440 | 0.424 | 0.235 | 0.265 | 0.225 | 0.281 | 0.202 | 0.208 | 0.344 | 0.334 | 0.271 | 0.360 | 0.419 | 0.414 | 0.343 | 0.326 | 0.216 |
NerfStudio | 0.330 | 0.261 | 0.336 | 0.311 | 0.452 | 0.392 | 0.158 | 0.190 | 0.139 | 0.245 | 0.179 | 0.174 | 0.249 | 0.261 | 0.215 | 0.302 | 0.338 | 0.465 | 0.251 | 0.261 | 0.167 |
Gaussian Opacity Fields | 0.194 | 0.107 | 0.168 | 0.152 | 0.443 | 0.234 | 0.084 | 0.158 | 0.092 | 0.181 | 0.104 | 0.102 | 0.142 | 0.164 | 0.140 | 0.187 | 0.208 | 0.346 | 0.162 | 0.140 | 0.099 |
Gaussian Splatting | 0.193 | 0.101 | 0.165 | 0.160 | 0.350 | 0.222 | 0.095 | 0.169 | 0.103 | 0.156 | 0.113 | 0.107 | 0.170 | 0.171 | 0.160 | 0.190 | 0.177 | 0.266 | 0.153 | 0.141 | 0.108 |
Mip-Splatting | 0.196 | 0.098 | 0.165 | 0.158 | 0.354 | 0.226 | 0.095 | 0.172 | 0.104 | 0.159 | 0.112 | 0.109 | 0.155 | 0.172 | 0.161 | 0.197 | 0.176 | 0.265 | 0.159 | 0.137 | 0.108 |
Zip-NeRF | 0.153 | 0.113 | 0.153 | 0.159 | 0.317 | 0.183 | 0.067 | 0.106 | 0.069 | 0.131 | 0.076 | 0.081 | 0.095 | 0.119 | 0.083 | 0.152 | 0.153 | 0.218 | 0.127 | 0.121 | 0.081 |