Computer vision on edge devices

• TUBCloud Link for finetuned model weights, presentation slides and sample videos
• Documentation

Set up

If you are on Windows, make sure you've set up your system as specified here.

Before getting started, we recommend you delete any Ultralytics configuration file that may already be on your system.

Linux:

rm -rf ~/.config/Ultralytics/

Windows:

In most cases, the config can be found somewhere in $env:USERPROFILE\AppData.

Training

We've provided four datasets that you can use out of the box to train a model. Just make sure you follow the steps in the respective README files and convert the annotations into the YOLO format using the respective convert.py script. If so, you can simply start a training run on one dataset like this:

python3 train.py datasets/roadsigns/config.yaml

You can also specify the number of epochs to train for by passing it as an argument to the --epochs or -e flag (default is 1).

python3 train.py -e 3 datasets/roadsigns/config.yaml

Export

For better inference performance on edge devices, you will want to export your model to ncnn format. We recommend using the Ultraltics CLI for that.

yolo export model=runs/detect/train/weights/best.pt format=ncnn

Important note: When exporting to ncnn for the first time, Ultralytics will download the ncnn dependency. This takes a very long time (about 15 minutes average).

Inference

The app.py script sets up a simple flask server that streams images to a port. Provide a path to the model you want to use and the video source, usually a video file. The example video used here can be downloaded from here

python3 app.py runs/detect/train/weights/best.pt mittel_1.mp4

If you want to use your webcam as input, simply pass 0 as a video source.