Look Out for New Spiders: Araneae Species OOD Detection Using Ensemble Pooling

The order Araneae, commonly known as spiders, comprises more than 53,000 described species. In the last decade, roughly 9,000 species have been added and discovered, and the pace of discovery shows no signs of slowing down. Species discovery can help in biodiversity evaluation. Existing spider species discovery methods do not commonly incorporate machine learning, which we believe has promising potential to enhance discovery outcomes. To that effect, we present a new method for out-of-distribution (OOD) detection based on Mahalanobis distance scoring using ConvNeXt as a feature embedding model. Our method, OspideR, creates a separate anomaly detection model for each class in the training dataset. At inference time, a separate anomaly score is generated by each model, and then the minimum score is used to determine the final output score of the model, known as the OspideR score. Performance was evaluated using a test set containing in-distribution and out-of-distribution spider images, distinct from the training images. The experimental results show that OspideR achieves better AUROC, F1, and accuracy compared to the other methods tested.