from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("frame.jpg", imgsz=1280, conf=0.35,
                        iou=0.45, agnostic_nms=True)

SAHI tiled inference (recommended for small/distant vessels)

python

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

model = AutoDetectionModel.from_pretrained(
    model_type="yolov8",
    model_path="best.pt",
    confidence_threshold=0.30,
    device="cuda:0",
)
result = get_sliced_prediction(
    "frame.jpg", model,
    slice_height=640, slice_width=640,
    overlap_height_ratio=0.2, overlap_width_ratio=0.2,
    perform_standard_pred=True,
    postprocess_type="NMS",
    postprocess_match_threshold=0.5,
)

Training recipe

Base: Best_Maritime_Custom.pt (warm-started through run1 → run2 → run4)
Architecture: YOLO26L (24.7M params, 86.1 GFLOPs)
Image size: 1280
Optimizer: AdamW, lr0=0.003, cosine schedule
Augmentations: mosaic 0.5, mixup 0.1, hsv/translate/scale/fliplr
Early stopping patience: 40

Limitations

Trained on a single coastal scene; generalization to other geographies untested
Class 0/3/4/5 underrepresented in training data
Best with imgsz=1280; smaller sizes hurt small-vessel recall

MuayThaiLegz/yolo26l-vessel-coastguard-v1

YOLO26L Vessel Coast-Guard v1

Classes

Validation metrics (run4 best, epoch 126)

Usage

SAHI tiled inference (recommended for small/distant vessels)

Training recipe

Limitations

MuayThaiLegz/
yolo26l-vessel-coastguard-v1