You can deploy this model and start earning money today!

alimkacar/
argus-maritime-yolov8s

Name: argus-maritime-yolov8s
Availability: InStock

Views

No views yet

Argus — Maritime Real-Time Situational Awareness

Argus — Maritime YOLOv8s

Ship / boat / buoy detection for the marine environment

GitHub

Detector for the Argus maritime situational-awareness system. A YOLOv8s model fine-tuned to detect vessels, boats, sailboats and buoys, trained by merging multiple maritime datasets under one unified class taxonomy and evaluated cross-dataset on data it never saw during training.

Classes

boat, sailboat, vessel, buoy · _{(taxonomy also reserves person, obstacle)}

Training

Base model: yolov8s.pt (Ultralytics) · imgsz 640
Data: two maritime datasets merged under a unified taxonomy (Marine Surveillance + SeaShips, Roboflow YOLO format), remapped with a per-source class map.
Held-out set: a third dataset (Ship Detection) reserved entirely for cross-dataset evaluation — never seen during training.

Results

Validation (in-distribution)

Cross-dataset test (unseen)

Metric	Value
mAP@0.5	0.971
mAP@0.5:0.95	0.736
precision	0.974
recall	0.954

Metric	Value
mAP@0.5	0.823
mAP@0.5:0.95	0.611
precision	0.949
recall	0.777

Per-class results

Validation — 254 images, 1322 instances

Class	mAP@0.5	mAP@0.5:0.95
boat	0.968	0.636
sailboat	0.933	0.775
vessel	0.990	0.781
buoy	0.993	0.753

Cross-dataset test — 4660 images, 5853 instances

Class	mAP@0.5	mAP@0.5:0.95
boat	0.983	0.719
sailboat	0.974	0.719
vessel	0.394	0.331
buoy	0.942	0.673

Analysis. In-distribution performance is strong (mAP@0.5 0.971). On a completely unseen dataset the model still reaches 0.823 — good generalisation. boat, sailboat and buoy transfer excellently (0.94–0.98), while vessel drops (0.39): "vessel" spans a very broad, heterogeneous set of ship types across datasets and the coarse _default class mapping used when merging lumps many of them together. Refining and balancing the vessel taxonomy is the clear next step.

Sample detections

_{Detections on held-out test images}

Usage

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download("alimkacar/argus-maritime-yolov8s", "best.pt")
model = YOLO(weights)

results = model.predict("image.jpg", conf=0.35)
results[0].show()

Intended use & limitations

Intended use: maritime object detection in inshore / harbour daytime imagery; research and educational use.
Limitations: trained on a small number of merged sources, mostly daytime and inshore; generalisation to other cameras, night/thermal conditions and open-sea scenes is not guaranteed. The vessel class is the weakest under cross-dataset conditions.
Safety: not to be used as the sole input for safety-critical navigation or collision-avoidance decisions.

License & attribution

Trained with Ultralytics YOLOv8 — AGPL-3.0.
Training data (Roboflow Universe): Marine Surveillance (CC BY 4.0), SeaShips, Ship Detection. Please attribute the datasets when reusing.
Code & full project: https://github.com/alimkacar/argus