A Purely Visual Re-ID Approach for Bumblebees (Bombus terrestris)

Abstract

Entomologists have widely applied re-identification techniques to better understand insects and their interaction with the environment. While humans can re-identify other humans and some mammals quite well, entomologists rely on gluing markers on insects to perform this task. This paper presents an approach for purely visual re-identification of bumblebees (Bombus terrestris) without the need to use markers. Non-invasive identification methods offer the possibility to observe the interaction of bumblebees with their environment without disturbance. Both a CNN model and a simple body shape model were used to investigate how they can be re-identified within a colony. The best-performing model, BumbleNet, correctly identifies more than two-thirds (CMC-1 score) of the individuals. Bumblebees are known for their substantial variations in body shape. To understand whether other features can also play a role in re-identification, different augmentations are applied during the training of BumbleNet. It was found that non-body-shape features increased the performance of BumbleNet by 25 percentage points (CMC-1 score). This also explains the observed superiority of the CNN-based BumbleNet compared to the BumbleShape model, that is solely based on body size parameters.