Direct Sparse Odometry with Continuous 3D Gaussian Maps for Indoor Environments

1. Related Papers

Direct Sparse Odometry with Continuous 3D Gaussian Maps for Indoor Environments, Deng J, Lang F, Yuan Z, Yang X.

paper link

2. Installation

	git clone [email protected]:JD-hust/gs-dso.git

2.1 Environment

We run the code on Ubantu 20.04 with CUDA 11.8.

2.2 Required Dependencies

2.2.1 Suitesparse

Install with

	sudo apt-get install libsuitesparse-dev libboost-all-dev

2.2.2 Eigen3

Eigen 3.2.8, Follow Eigen Installation.

2.2.3 OpenCV

OpenCV 2.4.9, Follow OpenCV Installation.

2.2.4 Pangolin

Pangolin, Follow Pangolin Installation.

2.2.5 ziplib

Install with

	sudo apt-get install zlib1g-dev
	cd dso/thirdparty
	tar -zxvf libzip-1.1.1.tar.gz
	cd libzip-1.1.1/
	./configure
	make
	sudo make install
	sudo cp lib/zipconf.h /usr/local/include/zipconf.h

2.2.6 Libtorch

Libtorch 2.5.1, follow Libtorch Installation.

2.3 Build

	cd gs-dso
	mkdir build
	cd build
	cmake ..
	make -j4

3. Usage

3.1 Dataset Format

	<sequence folder name>
		|____________rgb/
		|____________data.ply
		|____________transfroms.json
		|____________groundtruth.txt
		|____________associate.txt/rgb.txt

Please adjust your dataset file directory and format as described above.

3.2 Run

To process prior lidar map, run it with the following instruction:

	./opensplat --val-render ../output -i PATH/to/datasets/ -o ../output/ -n 20000 --keep-crs --sh-degree 1

and then run it with the following instruction for localization:

	./gsdso_dataset files=PATH/to/datasets calib=../calib/DATASETS/calib.txt result=../output/ mode=1 preset=0

We provide a pre-trained prior map and an example sequence in the link.There is the instruction to run on the sequence:

	./gsdso_dataset files=../example/S2 calib=../calib/ICL/calib.txt result=../example/output/ mode=1 preset=0

For more details on configuration parameters, see Direct Sparse Odometry and OpenSplat.

4. Acknowledgement

This work is implemented based on Direct Sparse Odometry, RGBD-DSO and OpenSplat. Thanks to J. Engel et al., who open source such excellent code for community.