YOFLO

Real-time object detection and visual Q&A using Microsoft Florence-2. Like YOLO, but with language understanding.

Install

pip install yoflo

Quick Start

# Download model and run object detection on webcam
python -m yoflo -dm -od

# Or use the daemon for repeated fast queries
python -m yoflo.daemon start --model large
python -m yoflo.daemon query --image photo.jpg --action detect
python -m yoflo.daemon query --image photo.jpg --question "Is there a cat?" --action ask

Features

Object Detection (-od)

Detect objects in video streams or images.

# Detect all objects
python -m yoflo -dm -od

# Filter to specific classes
python -m yoflo -dm -od person car dog

# From RTSP stream
python -m yoflo -dm -od -rtsp rtsp://192.168.1.100:554/stream

# From YouTube live stream
python -m yoflo -dm -od -yt "https://www.youtube.com/watch?v=..."

Yes/No Questions (-ph)

Ask binary questions about what the model sees.

python -m yoflo -dm -ph "Is the door open?"
python -m yoflo -dm -ph "Is anyone wearing a hard hat?"

Inference Chains (-ic)

Ask multiple questions per frame. Returns PASS if 2+ are true.

python -m yoflo -dm -ic "Is there a person?" "Are they seated?" "Are they at a desk?"

Headless Mode (-hl)

No GUI window. Use on servers or for better performance (~20% faster).

python -m yoflo -dm -od -hl

Screenshot on Detection (-ss)

Auto-save timestamped screenshots when objects are detected.

python -m yoflo -dm -od cat -ss

Logging (-lf)

Write detections to alerts.log.

python -m yoflo -dm -od -lf

Video Recording (-r)

Record video based on triggers.

# Record when objects detected
python -m yoflo -dm -od -r od

# Record when inference is "yes", stop on "no"
python -m yoflo -dm -ph "Is there motion?" -r infy

# Record when inference is "no", stop on "yes"
python -m yoflo -dm -ph "Is the area clear?" -r infn

Multi-Camera (-wi, -rtsp)

Monitor multiple feeds simultaneously.

# Multiple webcams
python -m yoflo -dm -od -wi 0 1 2

# Multiple RTSP streams
python -m yoflo -dm -od -rtsp rtsp://cam1/stream rtsp://cam2/stream

Inference Rate Control (-il, -ir)

# Limit to 5 inferences per second
python -m yoflo -dm -od -il 5

# Display inference rate
python -m yoflo -dm -od -ir

Pretty Print (-pp)

Formatted detection output instead of raw JSON.

python -m yoflo -dm -od -pp

4-Bit Quantization (-4bit)

Reduce VRAM usage with quantized model.

python -m yoflo -dm -od -4bit

PTZ Camera Control

Control pan/tilt/zoom cameras via HID (USB) or ONVIF (IP network).

HID PTZ (Logitech PTZ Pro, etc.):

pip install hid

# Manual PTZ control with keyboard
python -m yoflo -dm -od -ptz

# Auto-track an object class
python -m yoflo -dm -od person -ptz track -to person

ONVIF PTZ (Hikvision, Dahua, Axis, most IP cameras):

pip install onvif-zeep
# or: pip install yoflo[onvif]

# Connect to ONVIF camera and auto-track
python -m yoflo -dm -od person -onvif 192.168.1.100 -onvif-user admin -onvif-pass password -to person

# ONVIF will auto-detect the camera's RTSP stream if no other source specified

Daemon Mode

Keep the model loaded in memory for fast repeated queries. No 90-second reload each time.

# Start daemon (loads model once)
python -m yoflo.daemon start --model large

# Query instantly
python -m yoflo.daemon query --image photo.jpg --action detect
python -m yoflo.daemon query --image "https://youtube.com/watch?v=..." --action detect
python -m yoflo.daemon query --image photo.jpg --question "Do you see a car?" --action ask

# Check status
python -m yoflo.daemon status

# Stop daemon
python -m yoflo.daemon stop

Python client:

from yoflo.client import YofloClient

client = YofloClient()
detections = client.detect("image.jpg")
answer = client.ask("image.jpg", "Is there a person?")

Model Options

Use -dm to auto-download, or -mp /path/to/model for local models.

All Flags

Requirements

• Python 3.8+
• CUDA GPU with 8GB+ VRAM (16GB recommended for large model)
• Works on Windows, Linux, macOS

License

MIT