A tool-kit for Wildlife Individual Identification that provides a wide variety of pre-trained models for inference and fine-tuning.

Documentation · Report Bug · Request Feature

Our other projects

MegaDescriptor

Wildlife Datasets

Introduction

The wildlife-tools library offers a simple interface for various tasks in the Wildlife Re-Identification domain. It covers use cases such as training, feature extraction, similarity calculation, image retrieval, and classification. It complements the wildlife-datasets library, which acts as dataset repository. All datasets there can be used in combination with WildlifeDataset component, which serves for loading extracting images and image tensors other tasks.

More information can be found in Documentation

Installation

To install wildlife-tools, you can build it from scratch or use pre-build Pypi package.

Using Pypi

pip install wildlife-tools

Building from scratch

Clone the repository using git and install it.

git clone git@github.com:WildlifeDatasets/wildlife-tools.git

cd wildlife-tools
pip install -e .

Modules in the in the wildlife-tools

• The data module provides tools for creating instances of the WildlifeDataset.
• The train module offers tools for fine-tuning feature extractors on the WildlifeDataset.
• The features module provides tools for extracting features from the WildlifeDataset using various extractors.
• The similarity module provides tools for constructing a similarity matrix from query and database features.
• The inference module offers tools for creating predictions using the similarity matrix.

Relations between modules:

  graph TD;
      A[Data]-->|WildlifeDataset|B[Features]
      A-->|WildlifeDataset|C;
      C[Train]-->|finetuned extractor|B;
      B-->|query and database features|D[Similarity]
      D-->|similarity matrix|E[Inference]

Example

1. Create WildlifeDataset

Using metadata from wildlife-datasets, create WildlifeDataset object for the MacaqueFaces dataset.

from wildlife_datasets.datasets import MacaqueFaces
from wildlife_tools.data import WildlifeDataset
import torchvision.transforms as T

metadata = MacaqueFaces('datasets/MacaqueFaces')
transform = T.Compose([T.Resize([224, 224]), T.ToTensor(), T.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))])
dataset = WildlifeDataset(metadata.df, metadata.root, transform=transform)

Optionally, split metadata into subsets. In this example, query is first 100 images and rest are in database.

dataset_database = WildlifeDataset(metadata.df.iloc[100:,:], metadata.root, transform=transform)
dataset_query = WildlifeDataset(metadata.df.iloc[:100,:], metadata.root, transform=transform)

2. Extract features

Extract features using MegaDescriptor Tiny, downloaded from HuggingFace hub.

from wildlife_tools.features import DeepFeatures

name = 'hf-hub:BVRA/MegaDescriptor-T-224'
extractor = DeepFeatures(timm.create_model(name, num_classes=0, pretrained=True))
query, database = extractor(dataset_query), extractor(dataset_database)

3. Calculate similarity

Calculate cosine similarity between query and database deep features.

from wildlife_tools.similarity import CosineSimilarity

similarity_function = CosineSimilarity()
similarity = similarity_function(query, database)

4. Evaluate

Use the cosine similarity in nearest neigbour classifier and get predictions.

classifier = KnnClassifier(k=1, database_labels=dataset_database.labels_string)
predictions = classifier(similarity['cosine'])
accuracy = np.mean(dataset_database.labels_string == predictions)

Citation

If you like our package, please cite us.

@InProceedings{Cermak_2024_WACV,
    author    = {\v{C}erm\'ak, Vojt\v{e}ch and Picek, Luk\'a\v{s} and Adam, Luk\'a\v{s} and Papafitsoros, Kostas},
    title     = {{WildlifeDatasets: An Open-Source Toolkit for Animal Re-Identification}},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {5953-5963}
}