Could I directly use the envs of original 3DGS for running this repo?

[laliwang]

First Thanks sincerely for your excellent work and sharing! I just wonder whether there is difference in submodules between SAD-GS and original 3DGS? What's more, I noticed that you have achieved accurate reconstruction under a single frame input. Have you proceeded to train the entire replica scene, and how is the optimization speed and video memory usage?

[kungfrank]

The submodules between SAD-GS and original 3DGS is different here because I change the diff_gaussian_rasterization a bit for pose optimization, which is not used in the paper. Therefore you can switch my diff_gaussian_rasterization with original 3DGS diff_gaussian_rasterization without any effect on SAD-GS performance. Regarding the optimization speed, this implementation is in pytorch using autograd, so the training speed is sub-optimal (about 5 times slower than original 3DGS).