Detection-Tools

Realtime YOLOv8 detector for workshop hand tools. A custom model identifies 11 tools from still images or a webcam (OpenCV). The same weights are intended for later deployment on an NVIDIA Jetson.

Features

• YOLOv8n custom training with Ultralytics
• Image inference and live webcam demo
• 11 workshop tool classes
• Dataset in standard YOLO format (train / valid / test)

Detected classes

ID	Class
0	gergaji_besi
1	gergaji_kayu
2	gunting
3	gunting_plat
4	kunci_t
5	meteran
6	penggaris_siku
7	tang_rivet
8	tang_scun_hidrolik
9	tekiro_c_clamp
10	water_pass

The model only detects these classes. Other objects are ignored or may be mislabeled.

Project structure

Detection-Tools/
├── app/
│   └── webcam_test.py       # live camera demo
├── notebooks/
│   └── 01_train_test.ipynb  # setup, train, test
├── train/  valid/  test/    # images + YOLO labels
├── data.yaml
└── best.pt                  # trained weights (keep local if not in git)

Requirements

• Python 3.10+
• A webcam for the live demo

pip install ultralytics opencv-python

Quick start (webcam)

1. Place best.pt in the project root.
2. In app/webcam_test.py, set ROOT to this project folder.
3. Run:

python app/webcam_test.py

On Windows, if python is not found:

py -3.10 app/webcam_test.py

Press Q in the video window to quit.

Train

Set path: in data.yaml to this project folder, then:

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    device=0,  # GPU; use "cpu" if no NVIDIA GPU
)

Training on CPU is slow. Google Colab (T4 GPU) is recommended. After training, copy runs/.../weights/best.pt to the project root.

Test on images

from ultralytics import YOLO

model = YOLO("best.pt")
model.predict(source="test/images", save=True, conf=0.5)

NVIDIA Jetson

1. Train on a GPU machine (PC or Colab).
2. Copy best.pt to the Jetson.
3. Export TensorRT when ready:

from ultralytics import YOLO

YOLO("best.pt").export(format="engine", imgsz=640, half=True)

Notes

The current labeled set is small (~62 images). Accuracy is limited, especially when a tool is held in hand (most training photos show tools on a table). More photos per class — handheld shots, varied lighting, and empty-background negatives — will improve results more than changing inference code.

When exporting from Roboflow, avoid Stretch resize to 512×512; keep the original aspect ratio (letterbox).