HARNES YOLOv8 β 3-class (excavator / arm / person)
YOLO26n, 30 epochs, imgsz 1280, dataset v6 25k+ images. Training 5-6 Nov 2026 (Kaggle).
| file | size | mAP50 |
|---|---|---|
best_v8_3class.pt |
5.3 MiB | 0.843 (excavator 0.913, arm 0.837, person 0.778) |
Anything.pt |
5.3 MiB | alias |
Class order (v8, different from old Core ML prototype):
- 0 = excavator
- 1 = excavator_arm
- 2 = person
Inference defaults (server): conf 0.35, imgsz 1600, classes 0+2.
Usage:
from huggingface_hub import hf_hub_download
path = hf_hub_download("Akhnaf03/HARNES-YOLOv8", "best_v8_3class.pt")
# or
from ultralytics import YOLO
model = YOLO(path)
Also used by HARNES-Server: set DETECTION_MODEL_PATH or let it auto-download.