Towards Individual Grevy's Zebra Identification via Deep 3D Fitting and Metric Learning
This paper combines deep learning techniques for species detection, 3D model fitting, and metric learning in one pipeline to perform individual animal identification from photographs by exploiting unique coat patterns. This is the first work to attempt this and, compared to traditional 2D bounding box or segmentation based CNN identification pipelines, the approach provides effective and explicit view-point normalisation and allows for a straight forward visualisation of the learned biometric population space. Note that due to the use of metric learning the pipeline is also readily applicable to open set and zero shot re-identification scenarios. We apply the proposed approach to individual Grevy's zebra (Equus grevyi) identification and show in a small study on the SMALST dataset that the use of 3D model fitting can indeed benefit performance. In particular, back-projected textures from 3D fitted models improve identification accuracy from 48.0% to 56.8% compared to 2D bounding box approaches for the dataset. Whilst the study is far too small accurately to estimate the full performance potential achievable in larger-scale real-world application settings and in comparisons against polished tools, our work lays the conceptual and practical foundations for a next step in animal biometrics towards deep metric learning driven, fully 3D-aware animal identification in open population settings. We publish network weights and relevant facilitating source code with this paper for full reproducibility and as inspiration for further research.