petrify. Freeze the world. Keep the clock ticking.

Open Source · Apache 2.0 · Built by starside.io

Freeze the world. Keep the clock ticking.

An open-source computer vision pipeline that removes people from security footage while preserving the live timestamp. Frame-perfect, zero-config, fully local.

See it in action GitHub

Scroll to explore

Live Demo

See the difference

Real security footage. Real results. Left side is the original recording. Right side is petrify's output. The timestamp keeps running. The people are gone.

📍 Driveway / Outdoor Camera

◀ Original │ petrify ▶

Driveway camera: person removed, timestamp preserved

Person enters from street. petrify holds the empty driveway frame while the burned-in clock continues ticking in real-time. Detected via MOG2 motion + HOG person shape.

📍 School Corridor / Indoor Camera

◀ Original │ petrify ▶

School corridor: standing person removed, timestamp preserved

Person walks then stands still. The static-diff detector catches them even after MOG2 adapts them into the background. The corridor looks perpetually empty.

Under the hood

Three detectors. One clean result.

A single detector is never enough. petrify chains three complementary signals so every subject is caught. fast movers, standing-still loiterers, and distant figures barely larger than noise.

→

🌫️

MOG2 Background Subtractor

Gaussian mixture model trained on the first 90 frames. Catches anything moving against the learned background, the primary detection net.

OpenCV MOG2

→

🧍

HOG Person Detector

Histogram of Oriented Gradients shape detector. Catches partial figures and distant silhouettes that produce too little motion for MOG2 to flag.

cv2.HOGDescriptor

🔒

Static Background Differ

Locked reference frame diff that never adapts. Finds subjects that stood still long enough for MOG2 to absorb them. The failsafe that never forgets.

BackgroundDiffer

Pass 1 Background Extraction

Scans every frame to identify the cleanest background window. The longest uninterrupted stretch with no detected motion. That window is median-composited into the locked reference frame that persists for the entire render.

Pass 2 Composite Rendering

Every frame is classified as person or not. Person frames swap in the locked background. The timestamp ROI is auto-detected via Laplacian edge scanning and composited live onto each frozen frame so the clock never stops.

Features

Everything you need. Nothing you don't.

Zero-config by default. Every parameter tunable when defaults aren't enough.

🔍

Auto ROI Detection

Automatically locates the timestamp bounding box using a two-pass frequency map + Laplacian edge scan. No manual coordinates needed.

🧠

Three-Detector Pipeline

MOG2 + HOG + static BackgroundDiffer run in sequence. Each catches what the others miss. No false negatives, minimal false positives.

🕐

Live Timestamp Overlay

The burned-in security clock is cropped from each original frame and composited onto every frozen output frame. Time is always authentic.

⚡

FFmpeg-Accelerated Output

Raw frames are written to a pipe and re-encoded by FFmpeg with H.264, preserving the original audio and container metadata.

🔧

Fully Tunable CLI

Every threshold, scan window, detector weight, and output codec is exposed as a flag. Start with zero args and tune from there.

🖥️

100% Local. Zero Cloud.

Runs entirely on your machine. No API keys, no upload, no data leaving your environment. Privacy by default, not by policy.

Get started

Up and running in 60 seconds.

Install once, run anywhere. petrify puts itself on your PATH via pipx. No virtualenv juggling.

terminal

# 1. Install system FFmpeg
$ brew install ffmpeg         # macOS
$ sudo apt install ffmpeg    # Linux

# 2. Install petrify (adds to PATH)
$ pipx install petrify

# Or from source
$ git clone https://github.com/starside-io/petrify
$ cd petrify && pipx install -e .

usage

# Zero-config. Auto-detects everything
$ petrify input.mp4 output.mp4

# Explicit ROI (faster, skips scan)
$ petrify input.mp4 output.mp4 --roi 0 55 700 60

# Higher sensitivity
$ petrify input.mp4 output.mp4 \
    --threshold 200 \
    --bg-diff-area-threshold 100

# Disable auto-ROI scan
$ petrify input.mp4 output.mp4 --no-roi

Zero config. Real results.

Run petrify input.mp4 output.mp4 and walk away. petrify will auto-detect the timestamp position, build the background reference, and produce a clean output. No flags required.

Read the full docs ↗

██████╗ ███████╗████████╗██████╗ ██╗███████╗██╗ ██╗ Freeze the world. Keep the clock ticking. Made with ❤️ by starside.io · https://starside.io → ROI auto-detect scanning 60 frames… found timestamp x=0 y=55 w=700 h=60 → Pass 1 background building reference… locked window frames 42–312 → Pass 2 render compositing 1847 frames… ✓ Done output.mp4 written

Performance

Benchmarks and requirements.

Tested on real security footage. No GPU required. Everything runs on CPU via OpenCV.

Benchmark MacBook Pro · M2 Pro · 16 GB

macOS 26.3 · Python 3.14 · FFmpeg 8.1 · opencv-python 4.13 · numpy 2.4

Video	Resolution	FPS	Duration	Wall time	Speed
camera.mp4	1920x1080	25	46.8s	45.7s	1.0x real-time
camera2.mp4	640x480	6	109.2s	4.95s	22x real-time

Requirements Minimum to run

Python 3.10+

opencv-python-headless 4.8+

numpy 1.24+

FFmpeg any recent

RAM 2 GB min · 4 GB for 1080p

GPU not required

Use cases

Any fixed camera. Any subject.

If there's a locked-down camera and a burned-in clock, petrify can strip the subjects while preserving the timeline.

🔏

Privacy compliance

removes Identifiable individuals

keeps Scene context · verified timeline

📦

Footage compression

removes Sparse motion (bird, leaf, flag)

keeps Static background · clean stream

📐

Space & layout analysis

removes People occluding floor plans

keeps Clean geometry · heatmap-ready

⚠️

Hazard detection

removes Humans masking static hazards

keeps Spills, obstacles, trip points

🤖

AI training data

removes Subject identity from scenes

keeps Environment · ground truth labels

🌐

Any fixed camera

removes All foreground subjects

keeps Clean background · live timestamp

ℹ️ Educational & research use. petrify is an open-source tool intended for lawful, authorised applications only. Always ensure you have the legal right to process any footage, including compliance with applicable privacy regulations (PDPA, GDPR, CCPA, and local equivalents). Do not process footage of individuals without proper authorisation, consent, or a lawful basis.