Skip to main content

CLI Use (YOLO on Jetson)

This section explains how to use the YOLO command-line interface (CLI) on NVIDIA Jetson devices for model inference and testing.

1. Download Source Code​

Clone the YOLO repository and enter the working directory:

git clone https://github.com/ultralytics/ultralytics.git
cd ultralytics

(Optional) Install dependencies:

python3 -m pip install --upgrade pip
pip install -r requirements.txt

2. Enable Optimal Performance on Jetson​

To achieve the best inference performance, configure Jetson for maximum performance.

2.1 Enable MAX Power Mode​

sudo nvpmodel -m 0

Verify the current power mode:

sudo nvpmodel -q

2.2 Enable Jetson Clocks​

Lock CPU, GPU, and memory clocks:

sudo jetson_clocks

Restore default clocks if needed:

sudo jetson_clocks --restore

3. YOLO CLI Prediction Examples​

3.1 Image Prediction​

yolo predict model=yolov8n.pt source=image.jpg device=0

3.2 Video Prediction​

yolo predict model=yolov8n.pt source=video.mp4 device=0

3.3 USB Camera Prediction​

yolo predict model=yolov8n.pt source=0 device=0
note

source=0 corresponds to /dev/video0.

3.4 CSI Camera Prediction (GStreamer)​

yolo predict model=yolov8n.pt source="nvarguscamerasrc ! video/x-raw(memory:NVMM), width=1280, height=720, framerate=30/1 ! nvvidconv ! video/x-raw, format=BGRx ! videoconvert ! video/x-raw, format=BGR ! appsink"

4. Output Results​

Prediction results are saved by default to:

runs/detect/predict/

This directory contains: - Annotated images or videos - Detection metadata

5. Verification​

Check YOLO environment status:

yolo checks

Expected output includes: - CUDA available - GPU detected - Torch installed correctly

Summary​

  • YOLO CLI enables rapid testing on Jetson
  • Supports image, video, USB, and CSI camera inputs
  • Use MAX power mode for best performance
  • Suitable for development and validation workflows

Maintained by HemiHex for Jetson-based advanced vision workflows.