recipe: opencv-contrib-python 5.0.0.93

[ndonkoHenri]

Adds opencv-contrib-python — OpenCV plus the opencv_contrib extra modules — for iOS and Android, as a twin of the already-landed opencv-python 5.0.0.93.

Requested by flet-dev/flet#4712

What it adds over opencv-python

The opencv_contrib modules: aruco (AR markers), face, tracking, ximgproc (edge-aware filters, superpixels), wechat_qrcode, img_hash, xfeatures2d (non-free off), dnn_superres, and more — the same importable cv2, just with these extra namespaces available.

Shape

The contrib sdist ships the identical opencv/ source tree (it only ADDS opencv_contrib/ and points the extra-modules path at it), so the recipe is a byte-for-byte twin of opencv-python: meta.yaml and mobile.patch mirror it, and the v5 patch (14 files) applies unchanged — verified by real-applying to the opencv-contrib-python-5.0.0.93 sdist. Same Flet 0.86 cv2 loader fix (find_spec('cv2.cv2') → ExtensionFileLoader('cv2', origin) + extract_packages: [cv2]), same iOS MLAS gate.

One contrib-only iOS detail: the rgbd module's hand-written pyopencv_linemod.hpp uses a bare Ptr<> that clashes with Apple's <MacTypes.h> under using namespace cv, so the iOS lane sets -DBUILD_opencv_rgbd=OFF. Android has no MacTypes.h and keeps rgbd. Restoring it on iOS is a 4-line cv::Ptr follow-up.

Consumer notes (usage & recommendations)

opencv-contrib-python is OpenCV with the opencv_contrib extra modules folded into the same import cv2 — aruco, face, tracking, ximgproc, wechat_qrcode, img_hash and more, on top of everything the base OpenCV gives you.

Install

numpy comes in automatically as an OpenCV dependency:

# pyproject.toml
dependencies = [
    "flet",
    "opencv-contrib-python",
]

[tool.flet.android]
extract_packages = ["cv2"]

Minimal app

Run a real cv2 pipeline and show the result. There's no GUI on a phone, so display images through Flet controls rather than cv2.imshow:

import flet as ft
import cv2
import numpy as np


def main(page: ft.Page):
    img = np.zeros((200, 200, 3), dtype=np.uint8)
    cv2.circle(img, (100, 100), 70, (0, 140, 255), -1)
    edges = cv2.Canny(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY), 50, 150)
    ok, png = cv2.imencode(".png", edges)

    page.add(
        ft.Text(
            f"cv2 {cv2.__version__}: {int((edges > 0).sum())} edge pixels",
            weight=ft.FontWeight.BOLD,
        ),
        ft.Image(src=png.tobytes(), width=200, height=200),  # src takes bytes
    )


ft.run(main)

For live/streaming frames (video, real-time filters, per-frame detection), ft.RawImage is a better fit than ft.Image — it takes a NumPy array directly (no per-frame PNG encode) and streams frames over a data channel at interactive rates. One caveat: OpenCV is BGR, RawImage expects RGB, so convert first:

raw = ft.RawImage(width=320, height=240)
page.add(raw)
for frame in your_frames():                        # BGR uint8 arrays
    await raw.render(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))   # numpy straight in

Recommendations

• Don't install both opencv-python/opencv-contrib-python (or their headless variants) — all provide the cv2 module and will collide. Pick one; use contrib only if you actually need the extra modules.
• cv2.rgbd / cv2.linemod are Android-only in this build (the iOS gate above). Everything else — aruco, face, ximgproc, wechat_qrcode, img_hash, dnn_superres, … — is on both platforms.
• No GUI calls — cv2.imshow/waitKey don't apply; render via ft.Image(src=<png bytes>) or ft.RawImage as shown.
• Heavy processing off the UI thread — wrap long pipelines in page.run_thread(...).