Skip to main content
WCS
Menu
Library
Library Catalog
eJournals & eBooks
WCS Research
Archives
Research Use
Finding Aids
Digital Collections
WCS History
WCS Research
Research Publications
Science Data
Services for WCS Researchers
Archives Shop
Bronx Zoo
Department of Tropical Research
Browse By Product
About Us
FAQs
Intern or Volunteer
Staff
Donate
Search WCS.org
Search
search
Popular Search Terms
WCS History
Library and Archives
Library and Archives Menu
Library
Archives
WCS Research
Archives Shop
About Us
Donate
en
fr
Title
Comparison of two individual identification algorithms for snow leopards (Panthera uncia) after automated detection
Author(s)
Bohnett, Eve;Holmberg, Jason;Faryabi, Sorosh Poya;An, Li;Ahmad, Bilal;Rashid, Wajid;Ostrowski, Stephane
Published
2023
Publisher
Ecological Informatics
Published Version DOI
https://doi.org/10.1016/j.ecoinf.2023.102214
Abstract
Photo-identification of individual snow leopards (Panthera uncia) is the primary data source for density estimation via capture-recapture statistical methods. To identify individual snow leopards in camera trap imagery, it is necessary to match individuals from a large number of images from multiple cameras and historical catalogues, which is both time-consuming and costly. The camouflaged snow leopards also make it difficult for machine learning to classify photos, as they blend in so well with the surrounding mountain environment, rendering applicable software solutions unavailable for the species. To potentially make snow leopard individual identification available via an artificial intelligence (AI) software interface, we first trained and evaluated image classification techniques for a convolutional neural network, pose invariant embeddings (PIE) (a triplet loss network), and compared the accuracy of PIE to that of the HotSpotter algorithm (a SIFT-based algorithm). Data were acquired from a curated library of free-ranging snow leopards taken in Afghanistan between 2012 and 2019 and from captive animals in zoos in Finland, Sweden, Germany, and the United States. We discovered several flaws in the initial PIE model, such as a small amount of background matching, that was addressed, albeit likely not fixed, using background subtraction (BGS) and left-right mirroring (LR) techniques which demonstrated reasonable accuracy (Rank 1: 74% Rank-5: 92%) comparable to the Hotspotter results (Rank 1: 74% Rank 2: 84%)The PIE BGS LR model, in conjunction with Hotspotter, yielded the following results: Rank-1: 85%, Rank-5: 95%, Rank-20: 99%. In general, our findings indicate that PIE BGS LR, in conjunction with HotSpotter, can classify snow leopards more accurately than using either algorithm alone.
Keywords
Background subtraction; Deep learning; Hotspotter; Individual identification; PIE v2; Snow leopards
Access Full Text
A full-text copy of this article may be available. Please email the
WCS Library
to request.
Back
PUB36074