WhaleDrone

WhaleDrone yolo26n-obb Model

This repo contains yolo26n-obb-best.pt, the best checkpoint from a YOLO26n oriented bounding-box (OBB) model fine-tuned on the WhaleDrone Los Cabos humpback whale UAV dataset.

Model

• Task: Oriented bounding-box object detection
• Base model: yolo26n-obb.pt
• Checkpoint: yolo26n-obb-best.pt
• Classes: whale, boat, dolphin
• Parameters: YOLO26n model variant

The filtered dataset contains 450 images and 997 labeled instances. It has 407 training images / 914 instances and 43 validation images / 83 instances. The dataset contains 858 whale instances, 139 boat instances, and no dolphin instances.

Training

Training was performed with Ultralytics using the following main settings:

• 100 epochs
• Image size: 1024
• Batch size: 16
• Optimizer: auto
• Pretrained initialization enabled
• Automatic mixed precision enabled
• Seed: 0
• Deterministic training enabled

The dataset configuration is available at dataset.yaml, with the complete dataset summary in yolo_dataset_filtered/README.md.

Validation performance

The best epoch was selected by validation mAP50-95 from the training run:

These metrics are aggregate validation results reported by Ultralytics for the OBB task. They should be interpreted with the class imbalance and the absence of dolphin examples in mind.

Inference

Install Ultralytics, then run inference with:

yolo obb predict \
  model=best.pt \
  source=/path/to/image-or-video \
  imgsz=1024

For Python:

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict(source="/path/to/image-or-video", imgsz=1024)

The model predicts oriented boxes for whales and boats in this filtered dataset. Because no dolphin instances were present during training or validation, dolphin predictions are not expected to be reliable.