# YOLOv8

> [name=江廷威 ([e-mail])]

## Tools used

- [Ultralytics](https://docs.ultralytics.com/#how-can-i-get-started-with-yolo-installation-and-setup)
- [Roboflow](https://roboflow.com/)
- [Kaggle](https://www.kaggle.com/) / [Colab](https://colab.google/)

## Install

reference: https://docs.ultralytics.com

pip install ultralytics

## Data collection

### Capture video

This is the code for recording mp4 video using opencv, note that we should set the correct ***height*** and ***width*** supported for the camera.
We are saving mp4 video because the [lableing tool](#Labeling) we use only support mp4 video or images.

import cv2

cap = cv2.VideoCapture(0)
cap.set(3,640) # width
cap.set(4,480) # height

fourcc = cv2.VideoWriter_fourcc(*'MP4V')
out = cv2.VideoWriter('output.mp4', fourcc, 20.0, (640,480)) # (width, height)

while(True):
    ret, frame = cap.read()
    out.write(frame)
    cv2.imshow('frame', frame)
    c = cv2.waitKey(1)
    if c & 0xFF == ord('q'):
        break

cap.release()
out.release()
cv2.destroyAllWindows()

### Labeling

There are many labeling tools out there, here we use https://roboflow.com/

1. Upload video or images


2. Assign labeling work to teammates


3. Label the ball


4. Export data


## Train

Training without GPU is ***extremely slow***. If you don't have a GPU, use kaggle or google colab.

Use this [jupyter notebook](https://github.com/NCTU-AUV/yolov8-training/blob/master/yolov8.ipynb) to train on kaggle and export to .onnx format

1. Upload a yolov8 (.zip file) dataset to kaggle.
2. Change the name of the dataset in the notebook.
3. Run all
4. Download `best.onnx`

## Predict

refer to: https://hackmd.io/5pV_GR3qSLqnGV9uZBOLpg#isaac_ros_yolov8