Deep learning for honeybees identification

Abstract

This work addresses the problem of animal re-identification; it focuses on re-identifying honeybees based on their appearance in videos.<br /> <br /> Our first contribution comprises three image datasets, which consist of single honeybee images extracted from 12 days of video and annotated with information about their identity on long-term (multiple days) and short-term (a few seconds) scales. The first dataset contains 9,381 images with 78 known identities that are used to evaluate subject-dependent identification. The second dataset contains 8,962 images associated with 181 known identities and is used to evaluate the long-term re-identification of individuals. The third dataset contains 109,654 images associated with 4,949 short-term tracks that provide multiple views of an individual suitable for self-supervised training and short-term evaluation.<br /> <br /> We evaluated the Identification and Re-identification of honey bees in several setups using Convolutional neural networks models. Identification was evaluated with the subject-dependent dataset. Re-identification was evaluated in parts short-term and long-term. Short-term re-identification was evaluated to study how the performance behaves in the function of time. Re-identification was evaluated in test setups that capture different difficulty levels: from the same hour to a different day.<br /> <br /> Our work shows that a model trained with enough short-term data can outperform a model trained with less long-term data, even on evaluation in long-term scenarios.<br /> <br /> We also evaluated re-identification in more practical setups such as on a database setup and track re-identification. The ablation studies showing the impact of the quantity of data used in training and the impact of augmentation, will guide the design of future systems for individual identification.