Dependency alignment with DeepLabCut: opencv-python-headless and NumPy 2

[deruyter92]

Context

DeepLabCut installs FMPose3D via the optional [fmpose3d] extra. We are preparing a DLC release that:

• removes DLC’s numpy<2 upper bound for the default (PyTorch) install (DeepLabCut#3499)
• declares OpenCV explicitly on the DLC base deps as opencv-python-headless (DeepLabCut#3508)

pyproject.toml currently specifies non-aligned dependencies:

• numpy>=1.18.5,<2.0
• opencv-python>=4.5.0 (also again under the viz extra)

When users install deeplabcut[fmpose3d], that can:

1. block NumPy 2 even though DLC’s core path is ready for it
2. pull both opencv-python and opencv-python-headless (DLC/albumentations/napari prefer headless; FMPose3D/imgaug prefer full). Both distribute cv2, so install order decides which build you get.

Related: #36 already proposes removing NumPy (and related) upper bounds for DLC compatibility.

Questions

1. Can FMPose3D switch to opencv-python-headless?
   DLC does not need GUI HighGUI (cv2.imshow etc.), and headless avoids libGL issues on servers/CI/Colab.
   Would it be acceptable for FMPose3D to depend on opencv-python-headless instead of opencv-python?
   If something in FMPose3D requires the full OpenCV build, please say what — then we should document the dual-install.

2. Is numpy<2 still required?
   I.e. is the <2.0 pin there because of a known runtime break under NumPy 2, or was it inherited for compatibility with older stacks / DLC?

[xiu-cs]

Thanks @deruyter92 !

1. Yes, we can switch to opencv-python-headless. Relax dependency pins: torch<2.11, opencv-python-headless, remove duplicate opencv from viz extra #38
2. No, we have removed it in Remove numpy upper bounds for upcoming DeepLabCut dependency updates #36