kamera-linux/german-bird-classifier

German Bird Classifier 🐦

A fine-tuned EfficientNet-B0 model for classifying 8 common German garden bird species with 100% validation accuracy.

Model Description

This model was trained using vogel-model-trainer, a toolkit for creating bird classification models from video footage.

Base Model: google/efficientnet-b0 (8.5M parameters)
Fine-tuned for: German garden birds classification
Training Framework: PyTorch + Hugging Face Transformers

Supported Species

The model can classify the following 8 bird species:

1. Blaumeise (Blue Tit) - Cyanistes caeruleus
2. Grünling (European Greenfinch) - Chloris chloris
3. Haussperling (House Sparrow) - Passer domesticus
4. Kernbeißer (Hawfinch) - Coccothraustes coccothraustes
5. Kleiber (Eurasian Nuthatch) - Sitta europaea
6. Kohlmeise (Great Tit) - Parus major
7. Rotkehlchen (European Robin) - Erithacus rubecula
8. Sumpfmeise (Marsh Tit) - Poecile palustris

Performance

Species	Validation Accuracy	Samples
Blaumeise	100.0%	5
Grünling	100.0%	5
Haussperling	100.0%	5
Kernbeißer	100.0%	5
Kleiber	100.0%	5
Kohlmeise	100.0%	5
Rotkehlchen	100.0%	5
Sumpfmeise	100.0%	5
Overall	100.0%	40

Usage

With vogel-model-trainer

# Install the toolkit
pip install vogel-model-trainer

# Extract birds from video using this classifier
vogel-trainer extract --folder ~/bird-data \
  --species-model kamera-linux/german-bird-classifier \
  --sample-rate 20 --skip-blurry --deduplicate \
  video.mp4

With Python

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch

# Load model and processor
model = AutoModelForImageClassification.from_pretrained("kamera-linux/german-bird-classifier")
processor = AutoImageProcessor.from_pretrained("kamera-linux/german-bird-classifier")

# Load and preprocess image
image = Image.open("bird.jpg")
inputs = processor(images=image, return_tensors="pt")

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
    predicted_class = logits.argmax(-1).item()

# Get species name
species = model.config.id2label[predicted_class]
print(f"Predicted species: {species}")

Training Details

• Training Date: November 13, 2025
• Optimizer: AdamW with cosine LR schedule
• Augmentation: Random rotation, affine, color jitter, gaussian blur
• Regularization: Weight decay 0.01, label smoothing 0.1
• Early Stopping: 7 epochs patience
• Input Size: 224x224 pixels
• Batch Size: 32
• Learning Rate: 2e-5

Dataset

The model was trained on custom-collected video footage of German garden birds using the vogel-model-trainer toolkit:

• Video sources: Garden bird feeders in Germany
• Extraction method: YOLO-based bird detection + quality filtering
• Data split: 80% training / 20% validation
• Preprocessing: Blur detection, deduplication, class balancing

Limitations

• Trained specifically on German garden birds - may not generalize to other species or regions
• Best performance on clear, well-lit images similar to training data
• May struggle with:
  • Juvenile birds with different plumage
  • Birds in flight or unusual poses
  • Heavy occlusion or poor lighting
  • Species not in the 8 training classes

Ethical Considerations

This model is intended for:

• Wildlife monitoring and conservation
• Educational purposes
• Citizen science projects
• Automated bird feeder cameras

Not recommended for:

• Commercial wildlife tracking without proper permits
• Any use that could harm bird populations

Citation

If you use this model, please cite:

@software{german_bird_classifier_2025,
  author = {Kamera Linux},
  title = {German Bird Classifier},
  year = {2025},
  url = {https://huggingface.co/kamera-linux/german-bird-classifier},
  note = {Trained with vogel-model-trainer: https://github.com/kamera-linux/vogel-model-trainer}
}

Links

• Training Toolkit: vogel-model-trainer on GitHub
• PyPI Package: vogel-model-trainer
• Model Repository: kamera-linux/german-bird-classifier

License

Apache 2.0 - See LICENSE file for details.

Acknowledgments

Built with:

• Hugging Face Transformers
• PyTorch
• Google EfficientNet
• Ultralytics YOLO