Wildlife Pronghorn

Dataset

Images sourced from iNaturalist (research-grade observations). Bounding boxes generated by MegaDetector v5a (confidence ≥ 0.15). Split 80 / 10 / 10 train / val / test.

Performance

Evaluated on the held-out validation set (best checkpoint).

Usage

from ultralytics import YOLO
model = YOLO("models/pronghorn/yolo26s_finetuned_pronghorn_by_J.Gong_uwyo_2026-05-28.pt")
results = model.predict("image.jpg", conf=0.25)

--- license: cc-by-4.0 library_name: ultralytics pipeline_tag: object-detection tags: - wildlife - yolo - yolo26 - object-detection - camera-trap - pronghorn --- # Model Card — Pronghorn (*Antilocapra americana*) Single-class detection model for Pronghorn, fine-tuned from the Ultralytics YOLO26s backbone (pretrained on COCO). **Model file:** `yolo26s_finetuned_pronghorn_by_J.Gong_uwyo_2026-05-28.pt` ## Training Details | Property | Value | |----------|-------| | Base model | yolo26s.pt (COCO pretrained, Ultralytics) | | Architecture | YOLO26s | | Input size | 640 × 640 | | Epochs | 150 | | Optimizer | MuSGD, lr=0.002, momentum=0.9 | | Augmentation | mosaic=1.0, degrees=10°, scale=0.5, fliplr=0.5, hsv_h/s/v | | Device | NVIDIA RTX 5000 Ada Generation (32 GB, CUDA 12.8) | | Training date | 2026-05-28 | | Author | Jian Gong, University of Wyoming | ## Dataset Images sourced from iNaturalist (research-grade observations). Bounding boxes generated by MegaDetector v5a (confidence ≥ 0.15). Split 80 / 10 / 10 train / val / test. | Split | Images | |-------|-------:| | train | 215 | | val | 26 | | test | 28 | ## Performance Evaluated on the held-out validation set (best checkpoint). | Metric | Value | |--------|------:| | mAP50 | 0.8258 | | mAP50-95 | 0.6186 | ## Usage ```python from ultralytics import YOLO model = YOLO("models/pronghorn/yolo26s_finetuned_pronghorn_by_J.Gong_uwyo_2026-05-28.pt") results = model.predict("image.jpg", conf=0.25) ```