Quickstart — VigilCV

VigilCV is an ultra-fast, CPU-first data quality auditor and distribution drift sentinel for production Computer Vision pipelines.

Installation

pip install vigilcv

VigilCV requires Python 3.10+ and has no GPU dependencies. All computations run on CPU using NumPy and SciPy.

5-Minute Quickstart

1. Inspect a Single Image

from vigilcv import VisionSentinel

sentinel = VisionSentinel(
    blur_threshold=100.0,   # Laplacian variance minimum
    min_entropy=3.0,        # Shannon entropy minimum (bits)
    max_underexposure_ratio=0.20,
    max_overexposure_ratio=0.20,
    raise_on_fail=True,
)

metrics = sentinel.guard("path/to/image.jpg")
print(f"Blur score: {metrics.blur_score:.1f}")
print(f"Entropy:    {metrics.shannon_entropy:.2f} bits")
print(f"Passed:     {metrics.passed}")

2. Audit a Directory

from vigilcv import BatchAuditor

auditor = BatchAuditor(directory="dataset/train/", workers=8)
summary = auditor.run()

print(f"Total:     {summary.total_images}")
print(f"Valid:     {summary.valid_images}")
print(f"Corrupted: {summary.corrupted_count}")
print(f"Blurred:   {summary.blurred_count}")
print(f"Speed:     {summary.throughput_fps:.1f} FPS")

3. Use the CLI

# Inspect a single image
vigilcv inspect image.jpg

Audit a directory
vigilcv audit dataset/train/ --workers 8 --report report.html

Record a drift baseline
vigilcv baseline dataset/train/ --output baseline.pkl

Detect drift in new data
vigilcv drift dataset/new/ --baseline baseline.pkl

Core Dependencies

What's Next?

Core Concepts — Why signal heuristics beat gradient descent for pre-flight checks

Heuristics — The mathematical formulation of every quality gate

API Reference — Full VisionSentinel and BatchAuditor contracts