GS-W

Official GS-W implementation - 3DGS modified to handle appearance changes and transient objects. A reference view used to provide appearance conditioning. Note, that the method uses huge appearance embeddings (per-Gaussian) and appearance modeling has a large memory footprint.

Web: https://eastbeanzhang.github.io/GS-W/
Paper: Gaussian in the Wild: 3D Gaussian Splatting for Unconstrained Image Collections
Authors: Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, Haoqian Wang

Photo Tourism

Photo Tourism is a dataset of images of famous landmarks, such as the Sacre Coeur, the Trevi Fountain, and the Brandenburg Gate. The images were captured by tourist at different times of the day and year, images have varying lighting conditions and occlusions. The evaluation protocol is based on NeRF-W, where the image appearance embeddings are optimized on the left side of the image and the metrics are computed on the right side of the image.

Scene PSNR SSIM LPIPS Time GPU mem.
Sacre Coeur 19.73
Paper's PSNR: 23.24

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.824
Paper's SSIM: 0.8632

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.210
Paper's LPIPS: 0.13

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 1h 8m 56s 18.80 GB
Trevi Fountain 20.06
Paper's PSNR: 22.91

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.724
Paper's SSIM: 0.8014

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.272
Paper's LPIPS: 0.1563

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 1h 21m 34s 27.92 GB
Brandenburg Gate 24.35
Paper's PSNR: 27.96

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.903
Paper's SSIM: 0.9319

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.155
Paper's LPIPS: 0.0862

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 1h 10m 60s 19.06 GB
Average 21.38
Paper's PSNR: 24.70

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.817
Paper's SSIM: 0.865

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 0.213
Paper's LPIPS: 0.124

The original paper reports metrics for test images where the appearance embedding is estimated from the full test image, not just the left half as in the official evaluation protocol. The reported numbers are computed using the official evaluation protocol and are, therefore, lower than the numbers reported in the paper.

 1h 13m 50s 21.93 GB