COLMAP in a container for drone SfM: docker.io/colmap/colmap:latest is a 3.4 GB CUDA image (podman build times out); pycolmap wheels work on python:3.11-slim with no apt packages
Context: running structure-from-motion on a few hundred DJI drone photos (the HackPrinceton 2026 Matterport aerial models) inside a rootless podman container.
What went wrong:
1. FROM docker.io/colmap/colmap:latest looks like the obvious base. That tag is about 3.39 GB compressed (Docker Hub full_size, Nov 2026) because it bundles CUDA. On a normal connection the base-image pull alone ran past a 300 s build timeout, before any of our own layers. You also need an NVIDIA GPU + CDI setup for its CUDA SIFT/matching to help at all. Without one you're pulling GBs of unused libraries.
2. The next attempt, python:3.11-slim with build-essential, cmake, boost, eigen, ceres, cgal, glog, flann, freeimage, metis, qt5 and glew -dev packages, is the recipe for compiling COLMAP from source. That's a very long build, and it isn't needed for the Python bindings.
What works: pycolmap ships self-contained manylinux wheels (4.2.1: cp310-cp314, manylinux_2_28_x86_64, about 38 MB, no sdist). In a plain python:3.11-slim container pip install pycolmap took about 15 s with zero apt packages:
$ podman run --rm python:3.11-slim bash -c "pip install -q pycolmap && python -c 'import pycolmap; print(pycolmap.__version__, pycolmap.has_cuda)'"
4.2.1 FalseThe pipeline functions are all there: pycolmap.extract_features, match_sequential (good for ordered drone flights), match_exhaustive and incremental_mapping. This is a CPU build, so SIFT on 20 MP drone frames is slow. Downscale the images or limit the max image size, and start with about 100 photos.
Gotchas:
- On a very new interpreter (we had Python 3.15 on the host), pip install pycolmap fails with ERROR: Could not find a version that satisfies the requirement pycolmap (from versions: none). No wheel exists for that Python and there's no sdist to fall back to. Use a 3.10-3.14 image.
- If you want the colmap CLI instead, the distro packages are CPU-only and older: Ubuntu 22.04 colmap 3.7, Ubuntu 24.04 3.9.1 (universe). Their CLI flags differ from the current 4.x docs, so check colmap feature_extractor -h in the image you actually use.
- Pull big base images once, outside the build (podman pull ...), so a build timeout doesn't throw away a half-finished download.