This repo try to improve 3DGS so that it can reconstruct interior environment with good quality by using ARKit data.
The repository contains submodules, thus please check it out with
# SSH
git clone git@github.com:homee-ai/gaussian-splatting.git --recursiveor
# HTTPS
git clone https://github.com/homee-ai/gaussian-splatting.git --recursiveOur default, provided install method is based on Conda package and environment management:
cd gaussian-splatting/
conda env create --file environment.yml
conda activate homee_3DGS
# install hloc
git clone --recursive https://github.com/cvg/Hierarchical-Localization/
cd Hierarchical-Localization/
python -m pip install -e .Our preprocessor expects the following dataset structure in the source path location:
dataset
| ---- colmap
| ---- distort_images
| ---- sparse
| --- 0
| --- calibration.json
| --- distort_cameras.txt
| --- images.txt
| --- images_post.txt
| --- scene.obj (only for ARKit dataset)Run the preprocess script
bash prepare_dataset.sh <path to desire dataset colmap folder> [<method1> <method2> ...]After preprocessing, you will get the following dataset structure in the source path location:
dataset
| ---- colmap
| ---- distort_images
| ---- sparse
| ---- post
| ---- images
| ---- sparse
| ---- offline (if at least one pose-refining method is selected)
| --- method1
| --- final
| --- cameras.txt(bin)
| --- images.txt(bin)
| --- point3D.txt(bin)
| --- ...
| --- method2
| --- final
| --- cameras.txt(bin)
| --- images.txt(bin)
| --- point3D.txt(bin)
| --- ...
| --- ...
| ---- online
| --- cameras.txt
| --- images.txt
| --- point3D.txt
| ---- online_loop
| --- cameras.txt
| --- images.txt
| --- point3D.txt
| --- scene.obj (only for ARKit dataset)
| ---- <root>_nerfstudio (copy to directly train in nerfstudio)
| --- images
| --- colmap
| ---- arkit
| ---- 0
| --- cameras.txt
| --- images.txt
| --- point3D.txt
| ---- method1 (optional)
| ---- 0
| --- cameras.txt
| --- images.txt
| --- point3D.txt
| ---- method2 (optional)
| ---- ...
To run the optimizer, simply use
python train.py -s <path to dataset/colmap/post> -t <folder name under post/sparse> -m <path to desire output folder> --gs_type <gs or gs_mesh> --appearance_modeling- Download testing data from here.
- Copy testing data to 3DGS repo.
- Unzip testing data.
unzip data/tandt_db.zip - Train
python train.py -s data/db/playroom/ -t 0 -m output/playroom --gs_type gsThe output folder expect to be like the following dataset structure in the model path location:
output
| ---- point_cloud
| --- iteration_x
| --- point_cloud.ply
| --- ...
| --- ...
| ---- ...Download the trained scene (.ply) to local host and visulaized in polycam
-
Download Homee ARKit data from here.
-
To run the optimizer, simply use
python train.py -s <path to dataset> -t <folder name under sparse> -m <path to desire output folder> --gs_type <gs or gs_mesh> --appearance_modelingCommand Line Arguments for train.py
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
Path where the trained model should be stored (output/<random> by default).
Alternative subdirectory for COLMAP images (images by default).
We have two gs type, gs and gs_mesh, use gs_mesh if we have scene.obj file.
Enable appearance modeling or not.