#	Class	#	Class
1	cylinder_head	16	gear_box
2	alternator	17	chain
3	clutch_plate	18	disc
4	chain_sprocket	19	handlebar
5	cylinder_block	20	brake_clutch_lever
6	piston	21	brake_clutch_cable
7	crankshaft	22	brake_pad
8	camshaft	23	battery_terminal
9	spark_plug	24	fuse_box
10	air_fin	25	brake_light_switch
11	radiator	26	gear_lever
12	motorcycle_frame	27	brake_pedal
13	forks	28	carburetor
14	swingarm	29	fuel_tank
15	shock_absorbers	30	brake_oil_reservoir

Dataset (Private)

Total Images: 14,897
Train: 11,917
Validation: 2,980
Classes: 30
Avg/Class: ~400 train / 100 val

Training Details

Parameter	Value
Framework	Ultralytics YOLOv8
Base Model	yolov8n-cls.pt
Epochs	50
Image Size	224×224
Batch Size	32
Optimizer	AdamW (auto)
Hardware	Tesla T4
Training Time	~1.27 hours
Early Stopping	patience=15

Performance

Metric	Score
Top-1 Accuracy	94.5%
Top-5 Accuracy	99.2%

Sample Prediction

image 1/1 /content/NAR1018.jpg: 224x224 carburetor 0.98, alternator 0.01, shock_absorbers 0.00, air_fin 0.00, piston 0.00, 11.5ms Speed: 16.3ms preprocess, 11.5ms inference, 0.1ms postprocess per image at shape (1, 3, 224, 224)

Model Analysis

Confusion Matrix Insights

Strong diagonal dominance → excellent class separation
Minor confusion observed in:
- clutch_plate vs chain_sprocket
- piston vs cylinder_block
Overall low misclassification rate

Download and Usage

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Download model
model_path = hf_hub_download(
    repo_id="AswinG5/moto-parts-30cls",
    filename="motopartscls.pt"
)

# Load model
model = YOLO(model_path)

# Predict
results = model("your-image-file.jpg")[0]

Use Cases

Spare parts identification
Garage automation
Inventory systems
E-commerce tagging

Limitations

Lower performance on visually similar parts
Classes with fewer samples:
- clutch_plate
- chain_sprocket

Dataset Disclaimer

Dataset is private and not distributed.

Contact

My Linkedin https://www.linkedin.com/in/aswin-g3/