YOLO26 Real-Time Object Detection

A unified application for text-prompted object detection and segmentation using multiple backends:

• YOLOE-26L-Seg for fast real-time inference
• YOLO-World for open-vocabulary detection via Ultralytics
• Grounding DINO for high-accuracy zero-shot detection
• SAM3 for semantic segmentation

You can run the project as either:

• Web UI (app.py, Streamlit)
• CLI (cli.py, OpenCV window or headless mode)

Features

• Text-prompted detection (--prompt)
• Multiple input modes: image, video, webcam
• Backend selection: yolo, yolo-world, dino, sam
• Automatic hardware selection: MPS > CUDA > CPU
• Optional output saving for image/video workflows
• Headless processing for scripts/servers

Requirements

• Python >=3.11 (from pyproject.toml)
• macOS, Linux, or Windows
• Optional webcam (for live mode)
• Model files in project root:
  • yoloe-26l-seg.pt (YOLO backend)
  • sam3.pt (SAM backend)

Installation

Recommended (UV)

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone <repository-url>
cd realtime-detection-yolo26
uv sync --all-extras

Selective extras

uv sync --extra yolo
uv sync --extra dino --extra sam

Alternative (pip)

python -m venv venv
source venv/bin/activate
pip install .[yolo,dino,sam]

Quick Start

Web UI

uv run streamlit run app.py

Then open http://localhost:8501 and select backend + mode in the sidebar.

CLI

# YOLO (default)
uv run python cli.py image photo.jpg --prompt person

# Grounding DINO
uv run python cli.py image photo.jpg --type dino --prompt "person . car"

# YOLO-World
uv run python cli.py image photo.jpg --type yolo-world --prompt "person . car"

# SAM3
uv run python cli.py image photo.jpg --type sam --prompt "person"

CLI Reference

Syntax

uv run python cli.py <mode> [input] --type <yolo|yolo-world|dino|sam> --prompt "<classes>" [options]

Modes

• image - process one image
• video - process a video file
• webcam - process live camera frames

Common examples

# Save image output
uv run python cli.py image input.jpg --prompt car --output result.jpg

# Process video with frame skipping
uv run python cli.py video input.mp4 --prompt person --frame-skip 3 --output out.mp4

# Headless mode
uv run python cli.py video input.mp4 --prompt person --no-show --output out.mp4

# Webcam with a non-default camera
uv run python cli.py webcam --prompt person --camera-id 1

Options

Option	Description	Default
--type	Backend: yolo, yolo-world, dino, sam	yolo
--prompt	Prompt/classes string	None
--model	Custom model path	backend-specific
--output, -o	Output file path	None
--no-show	Disable OpenCV display	False
--device	auto, mps, cuda, cpu	auto
--camera-id	Webcam id	0
--frame-skip	Process every Nth video frame	1
--conf	Confidence (YOLO/SAM)	0.25
--box-threshold	Box threshold (DINO)	0.35
--text-threshold	Text threshold (DINO)	0.25

Interactive controls

• Webcam: press q to quit, s to save frame
• Video preview: press q to quit

Hardware Acceleration

The app auto-selects the best available device:

• MPS (Apple Silicon M1/M2/M3/M4)
• CUDA (NVIDIA GPU)
• CPU fallback

Verify detection setup:

uv run python test_device.py

Troubleshooting

Camera not found

• Grant camera permissions to terminal/IDE
• Try a different --camera-id (0, 1, 2...)

Model file missing

• Ensure yoloe-26l-seg.pt / sam3.pt exist in project root
• Or provide --model /path/to/model.pt
• YOLO-World defaults to Ultralytics yolov8s-world.pt, which downloads automatically if needed

MPS not available on macOS

• Requires macOS 12.3+ and Apple Silicon
• The app falls back to CPU automatically

Slow inference

• First inference is slower due to model warmup
• Use --frame-skip for large videos
• Reduce input resolution if needed

Project Structure

realtime-detection-yolo26/
|- app.py
|- cli.py
|- test_device.py
|- yoloe-26l-seg.pt
|- sam3.pt
|- README.md
|- ARCHITECTURE.MD
|- pyproject.toml
`- uv.lock

Documentation

• README.md - setup, quick start, and complete CLI usage
• ARCHITECTURE.MD - backend architecture, device/runtime model, and design details

License

See the Ultralytics License for model usage terms.