Multi-task learning for simultaneous pose estimation and re-identification of animals

Abstract

Monitoring animals plays a crucial role in animal conservation. Automated computer vision systems utilizing camera trap data are non-invasive alternatives to traditional tracking methods, providing a promising base for a scalable solution. This thesis builds on top of this computer vision-based animal tracking solution, called animal re-identification, and aims to find out whether learning the poses of animals enhances re-identification (re-ID) performance, in a multi-task learning setting. To investigate this, a Multi-Task Learning (MTL) model was developed with a modular architecture, allowing the auxiliary pose estimation branch to perform classification, regression, or be completely shut off, to establish the baseline. The model was evaluated under species-specific and closed-set conditions, meaning that it was trained to identify individuals of a single species, Eurasian lynxes, from a pool where all queried identities were already known to exist in the reference gallery. The experimental results indicate that using pose estimation as an auxiliary task does not yield higher re-ID accuracy compared to the baselines. However, the network successfully extracts both identity and pose information simultaneously, eliminating the computational need for a separate model.