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

• 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 evaluation module offers tools for creating and evaluating predictions using the similarity matrix.

  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[Evaluation]

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

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

metadata = StripeSpotter('datasets/StripeSpotter')
transform = T.Compose([T.Resize([224, 224]), T.ToTensor()])
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.

database = WildlifeDataset(metadata.df.iloc[100:,:], metadata.root, transform=transform)
query = WildlifeDataset(metadata.df.iloc[:100,:], metadata.root, transform=transform)

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(query), extractor(database)

Calculate cosine similarity between query and database deep features.

from wildlife_tools.similarity import CosineSimilarity

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

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

classifier = KnnClassifier(k=1)
predictions = classifier(similarity['cosine'])

pip install wildlife-tools

Following dependencies were used:

torch==2.0.1
pytorch-metric-learning==1.6.0
faiss-gpu==1.7.2
pycocotools==2.0.4
timm==0.9.2