Precise Horse Re-identification from a UAV Perspective
Loading...
Date
2025-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
For non-specialists, distinguishing similar animals can be challenging due to their subtle differences. This thesis evaluates machine learning’s ability to re-identify animals using UAV videos from rural settings. Videos were segmented to extract individual animal sequences, which formed the basis for feature extraction and re-identification. We used a pre-trained DINOv2 model to extract per-second features, which were aggregated and compared via similarity matrices computed with methods such as Average, Top20, and Maximum. Bipartite matching was applied to visualize identification performance and establish a baseline. Building on this, we fine-tuned the DINOv2 model with data augmentation strategies including random flipping and resizing. The fine-tuned model achieved perfect matching results in the baseline task. This work offers valuable insights for wildlife monitoring, animal behavior studies, and precision agriculture by enabling more accurate animal recognition. Future work will incorporate temporal dynamics and further refine the model to broaden its applicability across species and real-world scenarios.
Description
Keywords
Machine Learning, Computer Vision, Animal Monitoring, livestock managment, UAV Study