OpenCowID: Zero-Shot Visual Identification of Dairy Cows via Metric Learning

OpenCowID is a lightweight metric-learning framework for visual identification of dairy cows, designed to perform well even in zero-shot conditions. It uses stochastic texture synthesis to create diverse synthetic coat patterns, enabling robust embeddings without requiring large labeled datasets.

---

📦 Installation

pip install -r requirements.txt

🎨 Synthetic Coat Pattern Generation

To generate synthetic training data via stochastic texture synthesis, run:

python generate_data.py --directory_path /path/to/save/images

This script produces randomized cow-coat textures which serve as robust training signals for metric learning.

Data directory structure

data/
├── train_data/
│   ├── cow_001/
│   │   ├── img_001.jpg
│   │   ├── img_002.jpg
│   │   └── ...
│   ├── cow_002/
│   │   ├── img_001.jpg
│   │   ├── img_002.jpg
│   │   └── ...
│   ├── cow_003/
│   │   ├── img_001.jpg
│   │   └── ...
│   └── ...
│
├── val_and_gallery/
│   ├── gallery/                     # Reference images per cow ID
│   │   ├── cow_004/
│   │   │   ├── img_001.jpg
│   │   │   └── img_002.jpg
│   │   ├── cow_005/
│   │   │   └── img_001.jpg
│   │   └── ...
│   │
│   └── val/                         # Query images for validation
│       ├── cow_004/
│       │   ├── img_001.jpg
│       │   └── ...
│       ├── cow_005/
│       │   └── img_001.jpg
│       │   └── ...
│       └── ...
│
└── test_and_gallery/
    ├── gallery/                     # Reference images per cow ID
    │   ├── cow_006/
    │   │   ├── img_001.jpg
    │   │   └── img_002.jpg
    │   ├── cow_007/
    │   │   └── img_001.jpg
    │   └── ...
    │
    └── test/                        # Query images for testing
        ├── cow_006/
        │   ├── img_001.jpg
        │   └── ...
        ├── cow_007/
        │   └── img_001.jpg
        │   └── ...
        └── ...

🏋️‍♂️ Training

The training directory should follow a class-per-folder structure:

Run training with:

python train.py \
    --train_data_folder data/train_data \
    --val_gallery_folder data/val_and_gallery/gallery \
    --val_data_folder data/val_and_gallery/test

🔍 Testing

Prepare two directories:

• A gallery directory with one or more reference images per cow ID
• A test directory with query images.

Testing:

python infer.py \
    --model_path <path to model weights> \
    --train_data_folder data/test_and_gallery/gallery/ \
    --test_data_folder data/test_and_gallery/test/