Individual recognition of large carnivores

Description

The developed software allows individual identification of various species in the wild. The SW is based on the newly introduced MegaDescriptor and provides state-of-the-art performance on all publicly available datasets (see Table 1). Given the sequences of images and appropriate location data, the system provides the Top-3 visually similar individuals. Therefore, allows a more reliable identification, which consequentially enhances the efforts related to large carnivore protection.

Besides the MegaDescriptor, we also provide a submodule for the identification of Lynx lynx individuals. Compared to the original MegaDescriptor the LynxID module provides more reliable results. The module utilizes a MegaDescriptor fine-tuned on Lynx lynx data and can leverage sequential images of an individual to enhance the robustness of the result.

Lynx lynx Identification

We achieved more accurate results by fine-tuning MegaDescriptor with ArcFace loss on the Lynx lynx dataset. We utilized the Weights & Biases platform to conduct multiple sweeps over the hyperparameters. 

We used hyperparameter sweeps to find the optimal learning rate, learning rate scheduler, and parameters of the ArcFace loss function. We also conducted an ablation study on different augmentations and fine-tuned parameters of individual augmentations. The results of the parameter sweeps are available below.

Data Quality

Individual recognition tasks highly depend on the quality of input data. If a large enough part of the individual is not visible, identification is almost impossible. We observed improvement in results when we filtered the data and used only data in which individuals are captured from the side, as this angle typically highlights most of their distinctive features.

Furthermore, we improved accuracy by excluding individuals who were represented sparsely in the dataset, making it challenging to reliably estimate their feature representation.

If images of an individual are captured in sequence module can use this data for more robust predictions resulting in better accuracy. Additionally, this approach effectively handles predictions for lower-quality images.