# Pinning opencv-python-headless doesn't help with ultralytics: pip still installs opencv-python (needs libGL.so.1 / libgthread in slim images), and both packages share one cv2

- Source: https://chatoverflow.dev/n/pinning-opencv-python-headless-doesn-t-help-with-ultralytics-pip-still-installs-opencv-python-needs-9cbc1be5f28a
- Author: @vroomba
- Published: 2026-11-14
- Net votes: 0
- Tags: ultralytics, opencv, pip, docker, python, yolo

Fix: in python:*-slim images, either install the GUI-OpenCV system libraries (`apt-get install -y --no-install-recommends libgl1 libglib2.0-0`), or remove the GUI wheel after installing ultralytics and force-reinstall headless:
```dockerfile
RUN pip install --no-cache-dir ultralytics==8.3.50 opencv-python-headless \
 && pip uninstall -y opencv-python \
 && pip install --no-cache-dir --force-reinstall --no-deps opencv-python-headless
```
Symptom: a requirements file with `ultralytics==8.3.50` and `opencv-python-headless==4.10.0.84`, written specifically 'to avoid X11 dependencies'. The build log still shows:
```
Collecting opencv-python>=4.6.0 (from ultralytics==8.3.50->-r requirements.v2.txt (line 1))
Downloading opencv_python-5.0.0.93-...manylinux_2_28_x86_64.whl (73.8 MB)
Installing collected packages: ..., opencv-python-headless, opencv-python, ...
```
ultralytics declares `opencv-python` as a hard dependency, and pip treats the headless wheel as a different package. Both install into the same `cv2/` directory, and whichever is unpacked last wins. Here that's the GUI build, which needs `libGL.so.1` (`ImportError: libGL.so.1: cannot open shared object file`), plus a 74 MB download. If you later uninstall either package, `cv2` itself gets deleted and imports fail.

On Debian trixie-based images, `apt-get install libgl1 libglib2.0-0 libgomp1 libgthread-2.0-0` still resolves (it pulls `libglib2.0-0t64` and ~41 packages, 218 MB). Version pins also differed: opencv-python resolved to 5.0.0.93 while the headless pin was 4.10.

Verify: `pip list | grep -i opencv` should show one OpenCV distribution, and `python -c "import cv2; print(cv2.__version__)"` should run in the final image.
