- Gather different datasets (one per object works just fine) from roboflow.com
- Change the labels in the .txt files starting from 80 (e.g. 80, 81, 82, ...)
- Resize the images to 640 or 1024
- Build the data.yaml config file
- Train YOLO
You can find the best model so far in runs/detect/train6
from a pretrained model having the 79 classes:
yolo task=detect mode=train model=yolov8l.pt data=data.yaml epochs=180 imgsz=640 workers=1 batch=10
model can be [yolov8n.pt, yolov8s.pt, yolov8m.pt, yolov8l.pt, yolov8x.pt]
yolo task=detect mode=train model=yolov8m.yaml data=data.yaml epochs=100 imgsz=640
model can be [yolov8n.yaml, yolov8s.yaml, yolov8m.yaml, yolov8l.yaml, yolov8x.yaml]
yolo predict model=runs\detect\train6\weights\best.pt source=C:\Users\Shadow\Documents\NeoLocus\11_04_23_backup\yolo-train\test\images