Leopard Spotting
Projects | | Links: Report | Github
Leopard spotting facilitates the rangers and researchers track the leopard’s population and detect new pride members.
The model
We refer to leopard re-identification as the process of identifying a particular pride member upon a re-encounter. Unseen leopards, captured for the first time, are considered out-of-distribution members. Our model can re-identify and flag out of distribution pride members, facilitating researchers and rangers track their population in the wild. This model is the result of the capstone project for Summer 2022 at UC Berkeley. Colaborators: Carolina Arriaga, Madhu Hedge, Vish Pillai.
Check the website for more details: Leopard Spotting
Abstract
Vulnerable species, like leopards, living in protected areas face a rapid decline. Since conservation resources are limited, camera traps across the landscape help investigate animals’ space-use over large spatial distances. Researchers and computer vision engineers use deep learning technology to detect and identify animals based on images from camera traps. We implemented a deep learning pipeline of part detection and re-identification of known and out of distribution (OOD) leopards. We utilized 6k+ camera trap images with bounding box annotations for 431 African leopards from the Labeled Information Library of Alexandria: Biology and Conservation (LILA BC). The detection task involved YOLOv5 object detection with a performance of 80% mAP across leopard head, flank, and full body. Cropped leopard parts were sent through a ResNet-18 CNN embedding model and trained via a combination of contrastive and triplet loss function and softmax classifier. OOD detection was based on embedding distance from centroids and softmax values. For 64 leopard classes in the re-identification model, we were able to achieve an 80% top-5 accuracy and 67.5% top-1 accuracy on a test set with OOD rejection.
