A global model of bird detection in high resolution airborne images using computer vision

Bird Detection Datasets

Each dataset is organized into train and test splits, generally with 90% of images in train. Whereever possible the train/test split does not cross individual flights or locations. 

The general format is a csv with the columns: image_path, xmin, xmax, ymin, ymax, label

The coordinates relative to the image origin, there is no geographic projection in the images.

Bird Detection Models

 Using https://github.com/weecology/BirdDetector and the deepforest python package https://deepforest.readthedocs.io/. 

A single model for future use was trained using all training and test data together. (Bird.pt). 

Using the deepforest python package

```

from deepforest import main

import torch

m = main.deepforest()

m.model.load_state_dict(torch.load(<path to .pt>))

```

More information can found [biorxiv link].