Maritime_Custom v2 — YOLO26L Multi-Class Vessel Detection

Part of OmniSense / MediaSense — Connectivia Labs' industrial AI platform.

Upgrade from v1

• v1: single class vessel, 90.6% mAP@0.5, Singapore Maritime Dataset only
• v2: 10 vessel classes, ~16,500 images, 5 merged datasets

Performance

Metric	Value
mAP@0.5	72.1%
mAP@0.5:0.95	41.3%
Precision	65.8%
Recall	68.7%
Architecture	YOLO26L
Input size	1280×1280
Training date	2026-04-08

Per-class mAP@0.5

Class	mAP@0.5
boat	27.4%
sailboat	45.4%
fishing-vessel	61.8%
cargo-ship	46.9%
tanker	41.3%
ferry	41.3%
military-ship	50.7%
patrol-boat	28.9%
tugboat	37.5%
submarine	31.9%

Classes (10)

boat · sailboat · fishing-vessel · cargo-ship · tanker · ferry · military-ship · patrol-boat · tugboat · submarine

Datasets

Dataset	Images	License
ABOships (Zenodo)	9,880	CC BY
Military Ship Detection (Roboflow)	2,529	CC BY 4.0
Ship Detection 4-class (Roboflow)	~3,400	CC BY 4.0
Sea Vessels Engel (Roboflow)	698	CC BY 4.0
Aerial Maritime Solawetz (Roboflow)	74	MIT

Usage

from ultralytics import YOLO
model = YOLO('MuayThaiLegz/Maritime_Custom')
results = model('vessel_frame.jpg')

THREAT_CLASSES = {6, 7, 9}  # military-ship, patrol-boat, submarine
threats = [b for b in results[0].boxes if int(b.cls) in THREAT_CLASSES]

OmniSense · Pillar: MediaSense (Computer Vision) · Connectivia Labs 🔒