Automatic image-based re-identification of ringed seals

Abstract

Automated wildlife re-identification has attracted increasing attention in recent years as it provides a non-invasive tool to identify and track individual wild animals over time. Animal re-identification, together with access to a large amount of image material through camera traps and crowd-sourcing, provides novel possibilities for animal monitoring and conservation, in particular, when re-identifying individual animals from the images. The Saimaa ringed seal (Pusa hispida saimensis) is an endangered subspecies endemic to Lake Saimaa, Finland, and one of the few existing freshwater seal species. Ladoga ringed seals (Pusa hispida ladogensis) are a sister species of the Saimaa ringed seals that can only be found in Lake Ladoga. Ringed seals have permanent pelage patterns that are unique to each individual seal and can be used to identify any given member of the species. Their large variety of poses, further exacerbated by the deformable nature of seals together with varying appearance and low contrast between the ring pattern and the rest of the pelage, makes the task of re-identifying a ringed seal a challenge, providing a good benchmark to evaluate state-of-the-art re-identification methods.

In this study, the task of individual re-identification of the Saimaa and Ladoga ringed seals is solved by matching images based on animal pelage patterns. The general pipeline for the automatic processing of camera trap and handheld images of seals is proposed. The pipeline consists of three main steps: image preprocessing including seal segmentation, extraction of local pelage patterns and re-identification. Multiple approaches for each step are proposed and evaluated. Three metric learning-based frameworks for ringed seal re-identification: SaimaaID, NOvel Ringed seal re-identification by Pelage Pattern Aggregation (NORPPA) and LadogaID are developed. The extensive evaluation of the different methods are performed on a new and challenging Saimaa ringed seals re-identification dataset called SealID.

It is shown that a convolutional Neural Network-based method can be used to extract unique pelage patterns of the seals despite the low contrast of the pattern. Affine invariant patches are extracted using HessAffNet, decreasing the influence of different deformations. Different methods for patch embedding such as Triplet, SphereFace and HardNet Networks were applied and evaluated. The results of experiments show that Content-Based Image Retrieval (CBIR) algorithms which use feature aggregation achieve the best results. For the Ladoga ringed seals, an additional individual grouping step was developed. This step allows features to be aggregated from a sequence of images instead of a single image increasing the chances of correctly identifying each individual. Finally, discussion of the challenges encountered during the development of ringed seal re-identification methods is provided.