Cv Opencv

cv-opencv

Context: $ARGUMENTS

OpenCV 4.x via the Python cv2 binding. Python is the first-class surface — the C++, Java (javadoc), and JS (opencv.js) APIs are parallel and the same concepts port over. Canonical docs: https://docs.opencv.org/4.x/.

Quick start

• Inspect an image (shape/dtype/channels): → Step 2 (cv.py info)
• Grab N frames from a webcam / file: → Step 3 (cv.py capture)
• Detect faces with the modern YuNet model: → Step 4 (cv.py detect-faces)
• Track a bounding box across frames: → Step 5 (cv.py track)
• Run an ONNX model via the dnn module: → Step 6 (cv.py yolo)
• Calibrate a camera from a chessboard: → Step 7 (cv.py calibrate)
• Stitch a panorama: → Step 8 (cv.py stitch)

When to use

• User asks for face / object / landmark detection, optical flow, tracking, panorama, camera calibration, or any classical CV operation.
• Need to run an ONNX / TensorFlow / Caffe / Darknet model from Python without heavyweight frameworks (PyTorch, TF) — cv2.dnn is the lightweight path.
• Need to read/write video via FFmpeg, GStreamer, V4L2, AVFoundation, or MSMF with a uniform API (cv2.VideoCapture / cv2.VideoWriter).
• For pose/face landmark models trained by Google (MediaPipe Tasks), use the cv-mediapipe skill instead — it's a different ecosystem with its own Tasks API.

Step 1 — Install the right OpenCV wheel

Three PyPI wheels, mutually exclusive — picking the wrong one hides features or wastes install.

Install exactly one:

pip install opencv-contrib-python
# or via uv: uv pip install opencv-contrib-python

Never install both opencv-python and opencv-contrib-python — they conflict on cv2/*.so.

The helper scripts/cv.py declares opencv-contrib-python via PEP 723 so uv run handles it in an ephemeral venv.

Step 2 — Inspect images before processing

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py info path/to/image.png

Prints: shape (H, W, C), dtype (uint8 / uint16 / float32), channel count, min/max/mean per channel, and whether the file is readable by cv2.imread. If cv2.imread returns None, the file is either unreadable, the wrong format, or a non-ASCII path on Windows (use cv2.imdecode + np.fromfile in that case).

OpenCV's default channel order is BGR, not RGB. All operations assume BGR. Convert to RGB explicitly before handing to matplotlib / PIL / torchvision:

img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

Step 3 — Capture video frames (file / webcam / RTSP)

# webcam 0, 30 frames
uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py capture --src 0 --count 30 --out frames/
# RTSP stream
uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py capture --src 'rtsp://host/stream' --count 60 --out frames/
# video file
uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py capture --src in.mp4 --count 100 --out frames/

VideoCapture uses the first backend that accepts the URL. Force a backend with cv2.VideoCapture(src, cv2.CAP_FFMPEG | cv2.CAP_GSTREAMER | cv2.CAP_AVFOUNDATION | cv2.CAP_V4L2 | cv2.CAP_MSMF).

On macOS, webcam access requires Terminal to hold the camera permission (System Settings → Privacy & Security → Camera). On Linux, /dev/video0 must exist and be readable by your user.

Step 4 — Face detection (modern FaceDetectorYN)

Do NOT use Haar cascades for new work. cv2.FaceDetectorYN (YuNet) is the current recommendation — small (<1 MB), fast, accurate, works on any CPU.

# downloads model if not present in models/
uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py detect-faces path/to/image.jpg \
  --model models/face_detection_yunet_2023mar.onnx \
  --score-threshold 0.9

Download the YuNet ONNX from https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet. Outputs per-face bbox + 5 landmarks (eyes, nose, mouth corners).

For face recognition, pair with cv2.FaceRecognizerSF (SFace) — 512-d embeddings, cosine-distance matching.

Step 5 — Object tracking

The tracker API in 4.x moved out of cv2.legacy for the modern trackers:

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py track --src video.mp4 \
  --bbox "x,y,w,h" --tracker csrt --out tracked.mp4

Pick the bbox on the first frame; the script writes an annotated video.

Step 6 — DNN inference (ONNX / Caffe / Darknet / TF / Torch)

Readers (in order of popularity):

• cv2.dnn.readNetFromONNX(model.onnx) — most modern
• cv2.dnn.readNetFromTensorflow(frozen.pb, graph.pbtxt)
• cv2.dnn.readNetFromCaffe(deploy.prototxt, weights.caffemodel)
• cv2.dnn.readNetFromDarknet(cfg, weights) — YOLOv3/v4
• cv2.dnn.readNetFromTorch(model.t7) — legacy Torch
• cv2.dnn.readNet(path) — auto-detect by extension

Set backend + target AFTER loading:

net = cv2.dnn.readNetFromONNX("yolov8n.onnx")
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)  # or CUDA / INFERENCE_ENGINE / VKCOM
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CPU)       # or OPENCL / OPENCL_FP16 / CUDA / CUDA_FP16 / MYRIAD

Full backend × target compatibility matrix in references/dnn-backends.md.

Typical YOLO call (ONNX, letterbox-resize, NMS):

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py yolo \
  --model yolov8n.onnx \
  --input image.jpg \
  --conf 0.25 --iou 0.45 \
  --labels coco.names

The script uses cv2.dnn.blobFromImage + net.forward() + cv2.dnn.NMSBoxes.

Real OpenCV Python sample apps (upstream samples/dnn/) cover classification, detection, segmentation, openpose, text detection, face detect — they are the best reference for wiring new models.

Step 7 — Camera calibration

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py calibrate \
  --pattern chessboard --size 9x6 --square-mm 25 \
  --images 'calib/*.jpg' --out camera.yaml

Writes intrinsic matrix K, distortion coefficients, RMS reprojection error. 15+ images from different angles are recommended for a solid fit. For ArUco / ChArUco boards, use cv2.aruco.CharucoBoard + cv2.aruco.detectMarkers — better than chessboard at extreme angles.

Step 8 — Panorama stitching

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py stitch pano1.jpg pano2.jpg pano3.jpg --out pano.jpg

Uses cv2.Stitcher.create(). Modes: cv2.Stitcher.PANORAMA (default) or cv2.Stitcher.SCANS (flat docs). If it returns ERR_NEED_MORE_IMGS / ERR_HOMOGRAPHY_EST_FAIL, the overlap between input images is insufficient or feature matches are too sparse — feed more overlapping shots.

Binaries that ship with OpenCV

Real binaries (install via your distro or from source with cmake --build . --target install):

• opencv_version — prints 4.x.y. Sanity check.
• opencv_interactive-calibration — guided camera calibration (replaces the old calibrator).
• opencv_visualisation — single-sample visualiser for cascade classifiers.
• opencv_model_diagnostics — dnn model inspector.
• opencv_perf_*, opencv_test_* — only present when built with BUILD_TESTS=ON / BUILD_PERF_TESTS=ON.

Gotcha: opencv_traincascade and opencv_annotation were removed in OpenCV 4.x. They exist only on the 3.4 branch. If you need to train a new cascade, either use the 3.4 binary or retarget to modern DNN workflows (YOLO / SSD / YuNet).

Gotchas

• BGR by default. Every imread output is BGR. cv2.imshow expects BGR. Convert explicitly for RGB consumers.
• cv2.imread returns None silently on unreadable paths. Always check if img is None: after load.
• Windows + non-ASCII paths require cv2.imdecode(np.fromfile(path, np.uint8), cv2.IMREAD_COLOR) rather than cv2.imread.
• Tracker API split — prefer cv2.TrackerCSRT_create(), not cv2.legacy.TrackerCSRT_create(), unless you explicitly need the old API.
• cv2.dnn blobFromImage ordering matters — for most ONNX models you want swapRB=True (BGR→RGB) and scalefactor=1/255.0. Check the model card.
• HoughCircles is noisy — minDist, param1, param2 are extremely sensitive. Pre-blur (cv2.GaussianBlur(img, (9,9), 2)) before calling.
• VideoWriter codec strings are platform-dependent. mp4v writes a non-standard MP4; use avc1 for H.264 (requires FFmpeg backend). MJPG always works but bloats files.
• Coordinate order is (x, y) in most APIs but (row, col) == (y, x) when indexing numpy arrays. The canonical bug.
• cv2.findContours signature changed between OpenCV 3 → 4. In 4.x: contours, hierarchy = cv2.findContours(...) (2 returns). 3.x returned 3. Never unpack 3 in 4.x code.
• YuNet face detector REQUIRES setInputSize before detect (or it throws cryptic size errors). Pass (w, h) of the actual image.
• GPU CUDA backend needs OpenCV built WITH CUDA — wheels from PyPI are CPU-only. Build from source if DNN_TARGET_CUDA is required.
• opencv_traincascade is not in 4.x — see the note above.
• Haar cascades are legacy. They still work (cv2.CascadeClassifier) but produce many false positives vs YuNet. Use YuNet for any new face-detection work.
• ArUco moved to main repo in 4.7 — cv2.aruco is now bundled in opencv-python (no contrib needed). Older guides say contrib; check your version with cv2.__version__.

Examples

Example 1 — "Detect faces in a photo, draw boxes, save result"

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py detect-faces photo.jpg \
  --model models/face_detection_yunet_2023mar.onnx \
  --out photo_faces.jpg --draw

Download the model once:

curl -L -o models/face_detection_yunet_2023mar.onnx \
  'https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx'

Example 2 — "Track a car in a dashcam video"

# manually pick the bbox on frame 0 (x,y,w,h)
uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py track --src dashcam.mp4 \
  --bbox "640,360,120,80" --tracker csrt --out tracked.mp4

Example 3 — "Run YOLOv8 ONNX on a still image"

Export from ultralytics first: yolo export model=yolov8n.pt format=onnx. Then:

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py yolo \
  --model yolov8n.onnx --input street.jpg \
  --labels coco.names --out street_det.jpg --conf 0.3

Example 4 — "Calibrate a camera from a 9x6 chessboard"

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py calibrate \
  --pattern chessboard --size 9x6 --square-mm 25 \
  --images 'calib/IMG_*.jpg' --out camera.yaml

The YAML contains camera_matrix, dist_coeff, rms. Load with cv2.FileStorage.

Example 5 — "Stitch hand-held phone shots into a panorama"

uv run ${CLAUDE_SKILL_DIR}/scripts/cv.py stitch shot1.jpg shot2.jpg shot3.jpg shot4.jpg --out pano.jpg

Troubleshooting

cv2.error: (-215:Assertion failed) !_src.empty()

Cause: imread returned None — the file path is wrong, unreadable, or non-ASCII on Windows. Solution: Check os.path.exists(path) first; on Windows use cv2.imdecode(np.fromfile(path, np.uint8), cv2.IMREAD_COLOR).

AttributeError: module 'cv2' has no attribute 'FaceDetectorYN'

Cause: OpenCV < 4.5.4, or the module failed to compile. Solution: pip install -U opencv-contrib-python. Verify cv2.__version__ is >= 4.8.

cv2.error: OpenCV(4.x) ... cv::dnn::readNetFromONNX_ ... can't open file

Cause: model path wrong or file corrupted during download. Solution: check file size vs upstream; re-download via curl -L.

TrackerCSRT_create exists but track() returns false every frame

Cause: initial bbox empty or off-screen. Solution: verify (x, y, w, h) are inside the first frame and w > 0 and h > 0.

Webcam opens but returns all-black frames

Cause: OS-level camera permission not granted; wrong camera index; another app holds it. Solution: On macOS grant Terminal/iTerm Camera permission; try indices 0, 1, 2; close Zoom / FaceTime / OBS.

opencv_traincascade: command not found

Cause: Tool removed in 4.x. Solution: Use OpenCV 3.4 branch explicitly for cascade training, or move to modern DNN detectors (YuNet, YOLO, SSD).

Reference docs

• Read references/modules.md when you need a one-line description of every OpenCV module group (core, imgproc, imgcodecs, videoio, calib3d, features2d, objdetect, dnn, ml, photo, stitching, video, ximgproc…) plus the canonical URL for each.
• Read references/dnn-backends.md when choosing a cv2.dnn backend × target combination, especially for GPU / OpenVINO / Vulkan deployment.