This code is a non-official 2DGS implementation. The current code only implements the projection process in CUDA, including forward and backward. The depth distortion and normal regularization used in 2DGS will be updated soon. Relevant derivations can be found in 2DGS_code_supp.pdf, which is only available in Chinese.
This code is built based on official python demo https://colab.research.google.com/drive/1qoclD7HJ3-o0O1R8cvV3PxLhoDCMsH8W?usp=sharing#scrollTo=LTw-URSvR_it.
Ideally, in any environment configured with 3dgs, first run
pip install submodules/two-d-diff-gaussian-rasterization
then execute
python train.py --conf_path ./config/base.yaml
Here are some examples. A more comprehensive comparison will be released later. dtu scan65 (4000 steps)
Tanks and Temples (5000steps)
