Opencv

OpenCV Computer Vision AI Skill Guide

Overview & Engine Architecture

OpenCV (cv2) provides image/video IO, geometric and photometric transforms, classical feature detectors, and drawing utilities. Images are NumPy arrays in BGR order by default. Agents track color-space conversions explicitly, keep resize/crop parameters reproducible, release video captures, and push deep-learning detection to @pytorch / @huggingface-transformers when classical methods plateau.

imread / VideoCapture
      -> BGR ndarray
          -> cvtColor / resize / filter / threshold
          -> contours / features / write

When to use this skill

• Preprocessing images for ML models
• Classical detection (edges, contours, template match)
• Frame sampling and annotation from video

Operational directives

1. Remember cv2.imread returns BGR - convert before RGB-only libraries.
2. Check for None after reads; fail fast on missing paths.
3. Use cv2.IMREAD_COLOR / unchanged flags intentionally for alpha/bit depth.
4. Release VideoCapture / writers in finally blocks.
5. Do not hardcode absolute machine-specific GUI paths in headless servers (imshow needs a display).

Image preprocess example

import cv2
from pathlib import Path

path = Path("samples/part.jpg")
bgr = cv2.imread(str(path), cv2.IMREAD_COLOR)
if bgr is None:
    raise FileNotFoundError(path)

rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
resized = cv2.resize(bgr, (640, 480), interpolation=cv2.INTER_AREA)
gray = cv2.cvtColor(resized, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(blur, 50, 150)
cv2.imwrite("out/edges.png", edges)

Contours sketch

_, thresh = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
vis = resized.copy()
cv2.drawContours(vis, contours, -1, (0, 255, 0), 2)

Video frame sample

cap = cv2.VideoCapture("clips/line.mp4")
try:
    ok, frame = cap.read()
    if ok:
        cv2.imwrite("out/frame0.jpg", frame)
finally:
    cap.release()

Common failures

Best practices

• Keep calibration matrices and resize shapes with dataset versions.
• Normalize orientation via EXIF when feeding photos from phones.
• Use lossless PNG for intermediate masks; JPEG for previews.
• Combine with @pytorch for DNN modules (cv2.dnn) only when appropriate.

Limitations

• Not a full video editing suite (see dedicated video skills when present).
• Patent/licensing constraints may affect some algorithms in deployments.
• GPU modules (cuda) depend on build flags and drivers.

Related skills

• @pytorch / @huggingface-transformers - learned vision models
• @jupyter - interactive visualization of intermediates
• @ffmpeg - heavy video transcoding outside OpenCV