Binocular: A NABirds Species Classifier

This project provides a pretrained, species-level bird classifier for North American birds, built on the NABirds dataset. It uses a powerful DINOv2 Vision Transformer (ViT) backbone with a linear probe, allowing for accurate and efficient classification.

The model can be easily loaded and used for inference directly from the Hugging Face Hub.

Model Details

• Model Type: Vision Transformer (ViT) with a linear classification head.
• Backbone: DINOv2 (ViT-B/14).
• Dataset: NABirds - A large-scale dataset of North American birds with over 48,000 annotated images of 555 species.
• Resolution: Images are resized to 224x224 pixels.

How to Get Started with the Model

1. Installation

Install the package directly from PyPI:

pip install binocular-birds

2. Run Inference from the Command Line

After installation, you can use the predict_hf command-line tool to run inference on an image. The tool will automatically download the pretrained model from the Hugging Face Hub.

predict_hf --image "path/to/your/bird_image.jpg"

3. Run Inference in Python

You can also use the InferenceModel in your own Python scripts:

from Binocular.models.inference import InferenceModel
from PIL import Image

# This will download the model from the hub automatically
model = InferenceModel.from_pretrained(
    repo_id="jiujiuche/binocular",
    filename="artifacts/dinov2_vitb14_nabirds.pth"
)

# Open an image
img = Image.open("path/to/your/bird_image.jpg")

# Get predictions
predictions = model.predict(img, top_k=5)

# Print the results
for species, confidence in predictions:
    print(f"{species}: {confidence:.2%}")

Uses

This model is intended for classifying bird species in images. It can be used for:

• Academic research in ornithology and computer vision.
• Hobbyist applications for bird identification.
• As a baseline for more complex models.

Risks, Limitations and Biases

• The model is trained on the NABirds dataset and will perform best on images of North American birds. Its performance on out-of-distribution images (e.g., birds from other continents, low-quality images, or synthetic images) may be degraded.
• The dataset may contain biases related to geography, lighting conditions, and image quality, which could affect model performance.

Training

The full pipeline for training the model from scratch is available in the GitHub repository.

Dataset Setup

1. Download the NABirds dataset from https://dl.allaboutbirds.org/nabirds.
2. Extract the dataset files into the Binocular/datasets/nabirds/ directory within the cloned repository.

Training Command

The training process is managed through a main script that uses configuration files to define the training parameters.

# Example for training on an M2 Mac
python Binocular/scripts/train.py --config Binocular/configs/m2_dev.yaml

Training configurations for different hardware can be found in the Binocular/configs/ directory.

Citation

If you use this model in your research, please consider citing the original DINOv2 and NABirds papers, as well as this repository.

@misc{Binocular,
  author = {JiuJiuChe},
  title = {Binocular: A NABirds Species Classifier},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/JiuJiuChe/Binocular}}
}