You can deploy this model and start earning money today!

alimkacar/
argus-maritime-yolo11l

Name: argus-maritime-yolo11l
Availability: InStock

Views

No views yet

Argus v2 — Maritime YOLO11l

Argus — Maritime YOLO11l (v2)

Six-class maritime detection: vessels, boats, sailboats, buoys, people, obstacles

GitHub


🚢 vessel _0.692	🛶 boat _0.567	⛵ sailboat _0.583	🟡 buoy _0.585	🧍 person _0.511	🧱 obstacle _0.622

_{per-class mAP@0.5:0.95 on the full validation split}

Detector for the Argus maritime situational-awareness system. YOLO11l fine-tuned on four pooled maritime sources under one unified six-class taxonomy, with sequence-aware splits so no video clip spans two splits.

This is v2 and does not replace v1. v1 (alimkacar/argus-maritime-yolov8s) is a four-class YOLOv8s model measured on a different, easier split. The two numbers are not comparable; a same-split v1/v2 comparison is listed as pending below.

Classes

boat, sailboat, vessel, buoy, person, obstacle

Results

Full validation split: 6,875 images, 23,480 instances. Measured at 1280 px, the resolution the model was trained and selected at.

Metric	Value
mAP@0.5	0.854
mAP@0.5:0.95	0.593
precision	0.750
recall	0.805

Per class

Class	Instances	precision	recall	mAP@0.5	mAP@0.5:0.95
boat	1,794	0.592	0.856	0.842	0.567
sailboat	31	1.000	0.378	0.684	0.583
vessel	7,069	0.769	0.933	0.935	0.692
buoy	1,730	0.633	0.913	0.899	0.585
person	784	0.703	0.830	0.837	0.511
obstacle	12,072	0.803	0.919	0.925	0.622

Inference: 3.3 ms/image at 1280 px on an RTX PRO 6000 Blackwell. On Apple Silicon expect roughly 80–200 ms/image; no target-hardware benchmark is published yet.

Training


Base weights	`yolo11l.pt` (COCO pretrained)
Schedule	12 epochs at 1280 px, cosine LR, AMP + TF32
Early stopping	patience 4; best epoch 8, run ended at 12
Augmentation	mosaic on, closed for the last 4 epochs; no mixup, no multi-scale
Split sizes	27,267 train / 6,875 val / 4,335 test
Compute	~21 compute units (~2.3 h on one Colab G4 session)

A second stage was also trained (10 epochs at 1536 px, AdamW, mosaic off) and rejected by validation: see the analysis below. Both stage checkpoints were compared once on the same full validation split at 1536 px, and the higher mAP@0.5:0.95 checkpoint was published. The test split was never opened during training or selection.

Data

Four pooled sources under one taxonomy:

Source	Role	License
Marine Surveillance (Roboflow v9)	boats, sailboats, buoys	CC BY 4.0
Ship Detection (Roboflow v5)	cargo/fishing/tanker ships	CC BY 4.0
SeaShips7000 (Roboflow v1)	ship categories	CC BY 4.0
WaterScenes	`person`, `obstacle`, buoys, vessels	CC BY-NC-SA 4.0

Roboflow frames are grouped into video sequences and split 70/15/15 with a class-aware optimiser, so neighbouring frames never land in two splits. WaterScenes keeps its published train/val/test lists and is sampled every second frame (frame_stride: 2) because consecutive frames are near-duplicates. MODS is declared as a locked external benchmark and is never a training source.

No dataset images are redistributed with this model.

Analysis and limitations

Rare classes dominate the averaged metric. sailboat has 31 validation instances. The rejected 1536-px stage improved five of six classes (non-sailboat mean mAP@0.5:0.95 rose from 0.595 to 0.621) but stopped predicting sailboat entirely, and because mAP is a plain class average, losing one class of six cost 0.097 — far more than the 0.026 gained elsewhere. The likely cause is that stage 2 disabled mosaic augmentation, which had been inflating a rare class's per-batch exposure roughly fourfold. Treat the sailboat row as anecdotal; 31 instances cannot support a claim.

Class imbalance. obstacle (12,072) and vessel (7,069) dominate the loss against person (784) and sailboat (31).

Domain. Training data is daytime RGB from harbours, inland waterways and on-board cameras. Night, fog, heavy sea state, thermal/IR and radar are not represented. person labels derive largely from WaterScenes sailor annotations — distant, small people near boats, not close-up subjects.

Not yet measured: the locked test split, the same-split v1/v2 comparison, and target-hardware FPS/p95. Those artifacts are produced by argus.detection.evaluation_report and the video pipeline's --report-json, and are not part of this release.

Usage

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download("alimkacar/argus-maritime-yolo11l", "argus-v2.pt")
model = YOLO(weights)

# Predict at 1280 px: 640 loses buoys and people.
model.predict("harbour.jpg", imgsz=1280, conf=0.35)[0].show()

Licensing

Not legal advice; verify before any commercial use.

Ultralytics YOLO11 is AGPL-3.0, and weights derived from it inherit obligations when distributed.
WaterScenes training data is CC BY-NC-SA 4.0 — non-commercial and share-alike. Because it contributed to these weights, this card declares cc-by-nc-sa-4.0 and the model should be treated as research / non-commercial only.
The three Roboflow sources are CC BY 4.0; attribution is given above.

Citation

bibtex

@software{argus_maritime_yolo11l,
  title  = {Argus — Maritime YOLO11l (v2)},
  author = {Alim Kacar},
  url    = {https://github.com/alimkacar/argus},
  year   = {2026}
}