• Host machine with at least one NVIDIA GPU/CUDA support and installed drivers (to support dense reconstruction).
• Docker (for CUDA support 19.03+).

1. Check that Docker >=19.03 installed on your host machine:

   docker --version
   

2. Setup the NVIDIA driver and nvidia-toolkit on your host machine:

   For Ubuntu host machines: ./setup-ubuntu.sh For CentOS host machines: ./setup-centos.sh

3. Run the run script, using the full local path to your preferred local working directory (a folder with your input files/images, etc.):

   ./run.sh /path/where/your/working/folder/is
   

   This will put you in a directory (inside the Docker container) mounted to the local path you specified. Now you can run COLMAP binaries on your own inputs like this:

   colmap automatic_reconstructor --image_path ./images --workspace_path .
   

3-b. Alternatively, you can run the run-gui script, which will start the graphical user interface of COLMAP:

```
./run-gui.sh /path/where/your/working/folder/is
```

After completing steps 1-2, you can build the Docker image from scratch using the Dockerfile. First, update the CUDA and Ubuntu versions in Dockerfile lines 1-2 to match your system, then:

./build.sh
./run.sh /path/where/your/working/folder/is

Install an NVIDIA driver and NVIDIA container runtime:

sudo apt install ubuntu-drivers-common
sudo ubuntu-drivers autoinstall

If you failed to install the above, check the appropriate NVIDIA driver by yourself and install it:

ubuntu-drivers devices
e.g.
sudo apt install nvidia-driver-455