Foundation models for species-level categorization in animal re-identification pipelines

Abstract

Automated wildlife monitoring methods and camera traps have led to a rapid increase in image-based datasets. In animal re-identification, such datasets are commonly used to study individuals and populations.

Manual image categorization has been used for many years for wildlife species classification; however, analyzing large datasets is challenging due to the lack of efficiency and time consumption. This necessitates more efficient methods. The Segment Anything Model (SAM) version 3 is a pretrained foundational model that enables more hands-free and time-efficient classification. The primary species of interest in this work are Grévy’s zebra, savanna elephant, lion, plains zebra, and reticulated giraffe. The proposed pipeline first filters images into animal-positive and empty categories, compares the results against ground-truth labels, and then classifies the five target species in the positive animal images. Model performance is evaluated using accuracy metrics.

Since the model's accuracy is dominated by the majority class and the dataset contains significantly more animal-positive images than empty images, SAM v3 achieved up to 98.6\% accuracy in the first experiment for general animal detection. The low performance on the "False" class was related to classifying images containing printed animals or animal parts, such as skulls. This issue could be reduced by using more specific prompts. In the second experiment, the results show the model's strong ability to detect specific animal species. However, misclassification occurred between species with similar visual characteristics, such as different types of zebras.