KZHIwEI/mdpianimal-models

Edge-Deployed YOLO for Individual Identification of Cattle, Goats, and Mugger Crocodiles

PyTorch (Ultralytics) detection checkpoints behind the paper. 81 checkpoints: 9 detectors (YOLOv5/v8/v11 at n/s/m) × 3 datasets × 3 random seeds.

Layout

<dataset>/<model>/seed{0,1,2}.pt        dataset in {cattle, goat, crocodile}

yolov8m is the headline detector reported in the main text; the full set is the Appendix scale-sweep matrix.

Training

From COCO-pretrained weights, 500 epochs (early-stop patience 100), 640×640, batch 16, AdamW (auto). Strengthened augmentation: mosaic 1.0, HSV-H/S/V 0.02/0.7/0.4, rotation 90°, translate 0.1, scale 0.5, shear 10, perspective 0.0005, flipud 0.5, fliplr 0.0, dropout 0.2.

Results — mAP@[0.50:0.95] (3-seed mean±SD, held-out test set)

Model	Cattle	Goat	Crocodile
YOLOv5n	89.12±0.17	79.72±0.77	69.87±1.29
YOLOv5s	91.51±0.43	81.26±0.51	69.79±1.77
YOLOv5m	92.28±0.49	83.54±0.09	70.58±2.00
YOLOv8n	90.50±0.34	83.50±0.37	69.60±2.71
YOLOv8s	92.62±0.14	83.68±0.73	73.25±3.44
YOLOv8m	95.18±0.32	84.15±1.65	81.20±1.10
YOLO11n	90.48±0.76	85.73±0.20	75.78±2.87
YOLO11s	93.09±1.14	85.04±0.36	74.96±4.45
YOLO11m	93.15±1.10	84.92±0.66	69.09±4.47

Results — mAP@0.50 (3-seed mean±SD)

Model	Cattle	Goat	Crocodile
YOLOv5n	99.41±0.14	94.75±0.87	93.30±1.04
YOLOv5s	99.49±0.01	94.78±0.97	91.39±1.67
YOLOv5m	99.48±0.00	95.25±0.92	93.56±1.51
YOLOv8n	99.50±0.00	94.35±0.18	93.15±1.42
YOLOv8s	99.50±0.00	94.14±0.60	93.64±0.79
YOLOv8m	99.50±0.00	95.97±0.17	92.72±1.49
YOLO11n	99.50±0.00	95.14±0.82	93.35±0.46
YOLO11s	99.50±0.00	94.91±0.86	93.59±0.20
YOLO11m	99.46±0.05	94.73±1.23	93.27±0.29

Usage

from ultralytics import YOLO
m = YOLO("cattle/yolov8m/seed0.pt")
m.predict("image.jpg")

Datasets (640-px) are in the companion dataset repo KZHIwEI/mdpianimal.