This Dockerfile runs COLMAP with CUDA, allowing dense reconstruction. COLMAP supports Structure from Motion (SfM), Multi-View Stereo (MVS). It uses Ceres-Solver for bundle adjustment.

The image is built on CUDA 9.0, so the host machine must have CUDA >= 9.0.

Only the command line functions are available. The documentation is available at https://colmap.github.io/.

The NVIDIA runtime is required (nvidia-docker v2), instructions here: https://github.com/NVIDIA/nvidia-docker.

To build the image from the Dockerfile.

sudo docker build --no-cache  -t dgrnwd/colmap:gpu .

Connecting a host project directory to the container. Ensure the host project directory has an images sub directory.

sudo docker run --runtime=nvidia  \
-v /host/project:/project \
-it dgrnwd/colmap:gpu bash

Running the automatic reconstructor.

colmap automatic_reconstructor \
    --workspace_path /project \
    --image_path /project/images