Prompt-based segmentation for animal re-identification

Abstract

Animal re-identification is a crucial aspect of contemporary wildlife research and conservation efforts. It aims to facilitate non-invasive monitoring, wildlife population studies, and ecological research. Animal segmentation plays a vital role in animal re-identification, involving the delineation of animals within images or videos, often in challenging environments like the wild. The objective of this research was to leverage SEEM (Segment Everything Everywhere All at Once) for segmenting animals from the SealID dataset and the Bird dataset. SEEM segments animal images using various prompting techniques, such as textual descriptions, strokes, and reference images, to effectively guide the segmentation process. The developed pipeline was successfully evaluated against time, intersection over union (IoU), and accuracy. The datasets used for this evaluation were the SealID and bird datasets. Text-prompt achieved the highest performance with 83.25\% accuracy, followed by the example-prompt with 80\% accuracy, with the stroke-prompt coming in last having 69.30\% accuracy on the text-prompt approach for the SealID dataset.