Skip to content

README.md

Latest commit

 

History

History
231 lines (172 loc) · 5.31 KB

README.md

File metadata and controls

231 lines (172 loc) · 5.31 KB
Raw

Quickly Run YOLO26 on Jetson

This example runs YOLO26 object detection on Jetson devices with NVIDIA GPU acceleration using a lightweight custom Docker image.

Supported Devices

JetPack Device
5.x Jetson Xavier NX/AGX
6.x Jetson Orin Nano/NX
7.x reComputer J4012, Thor J6015

Quick Start

One-Command Run

Install jetson-examples:

pip3 install jetson-examples

Restart your reComputer:

sudo reboot

Run YOLO26 detection:

reComputer run yolo26

This will automatically:

  1. Build a lightweight Docker image (~500MB) with YOLO26 dependencies
  2. Mount host CUDA libraries (no duplication)
  3. Run inference on the bundled test video
  4. Save results to output/result.mp4

Custom Video

Run detection on your own video:

reComputer run yolo26 /path/to/your/video.mp4

Example:

# Run on a specific video file
reComputer run yolo26 ~/Videos/traffic.mp4

# Run on a video in the current directory
reComputer run yolo26 ./my_video.mp4

Results are saved to scripts/yolo26/output/result.mp4.

Model Export

Export to ONNX

# Inside the container or with ultralytics installed
yolo export model=yolo26n.pt format=onnx imgsz=640

Export to TensorRT

# For better performance on Jetson
yolo export model=yolo26n.pt format=engine imgsz=640 half=True

Export to NCNN (Lightweight)

# For edge deployment
yolo export model=yolo26n.pt format=ncnn imgsz=640

Running Without reComputer CLI

Build the Docker Image

cd /path/to/reComputer/scripts/yolo26
docker build -t yolo26:latest .

Run with Custom Video

docker run --rm \
    --runtime=nvidia \
    --network=host \
    -v /usr/local/cuda:/usr/local/cuda:ro \
    -v $(pwd)/output:/output \
    -v /path/to/video.mp4:/input/video.mp4:ro \
    -e NVIDIA_VISIBLE_DEVICES=all \
    -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \
    -e LD_LIBRARY_PATH=/usr/local/cuda/targets/sbsa-linux/lib:/usr/lib/aarch64-linux-gnu/nvidia \
    yolo26:latest \
    --source /input/video.mp4

Run with Camera

# Camera index 0 (default USB camera)
docker run --rm -it \
    --runtime=nvidia \
    --network=host \
    --device=/dev/video0:/dev/video0 \
    -v /usr/local/cuda:/usr/local/cuda:ro \
    -e NVIDIA_VISIBLE_DEVICES=all \
    -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \
    -e LD_LIBRARY_PATH=/usr/local/cuda/targets/sbsa-linux/lib:/usr/lib/aarch64-linux-gnu/nvidia \
    yolo26:latest \
    --source 0

Python Script Options

The detect.py script supports these arguments:

Argument Default Description
--source 0 Camera index (0) or video file path
--model yolo26n.pt Model file path (.pt, .onnx, .engine)
--conf 0.5 Confidence threshold (0.0 - 1.0)
--save auto Output path (auto: /output/result.mp4 if headless)
--imgsz 640 Inference size (pixels)

Examples

# Use custom model with high confidence
docker run --rm \
    --runtime=nvidia \
    --network=host \
    -v /usr/local/cuda:/usr/local/cuda:ro \
    -v $(pwd)/output:/output \
    -v /path/to/custom_model.pt:/app/custom_model.pt:ro \
    -v /path/to/video.mp4:/input/video.mp4:ro \
    -e NVIDIA_VISIBLE_DEVICES=all \
    -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \
    -e LD_LIBRARY_PATH=/usr/local/cuda/targets/sbsa-linux/lib:/usr/lib/aarch64-linux-gnu/nvidia \
    yolo26:latest \
    --source /input/video.mp4 \
    --model /app/custom_model.pt \
    --conf 0.7 \
    --imgsz 1280

Using Pre-built Ultralytics Images

If you prefer the official Ultralytics images:

# JetPack 7.x
t=ultralytics/ultralytics:8.4.54-nvidia-arm64
sudo docker pull $t
sudo docker run -it --ipc=host --runtime=nvidia $t

# JetPack 6.x
t=ultralytics/ultralytics:8.4.54-jetson-jetpack6
sudo docker pull $t
sudo docker run -it --ipc=host --runtime=nvidia $t

# JetPack 5.x
t=ultralytics/ultralytics:8.4.54-jetson-jetpack5
sudo docker pull $t
sudo docker run -it --ipc=host --runtime=nvidia $t

Performance

Typical performance on Jetson devices:

Device Model FPS Resolution
Thor J6015 YOLO26n 75 640x480
Orin Nano YOLO26n 45 640x480

Troubleshooting

CUDA Not Found

Ensure NVIDIA runtime is installed and configured:

# Check nvidia-smi
nvidia-smi

# Check Docker runtime
docker info | grep -i runtime

Video Cannot Open

  • Verify the video file exists and is accessible
  • Check file permissions
  • Ensure video codec is supported (MP4 with H.264 recommended)

Low FPS

  • Reduce input resolution: --imgsz 320
  • Use a smaller model (nano vs small)
  • Check GPU temperature: tegrastats

File Structure

yolo26/
├── Dockerfile          # Lightweight Docker image definition
├── detect.py           # Python detection script
├── run.sh              # Shell wrapper for docker run
├── config.yaml         # JetPack version and resource requirements
├── init.sh             # Environment initialization
├── test.mp4            # Sample video for testing
├── yolo26n.pt          # YOLO26 nano model
└── output/             # Detection results saved here

Reference