Analyzing MegaDetector V5 as a Binary Sorter for Ecological Camera Trap Images
Abstract/Description/Artist Statement
The rapid development of semi-automated workflows within Camera Trap detection has revolutionized the automatic detection and classification of animals. Building upon the foundational work of the MegaDetector object detection model, this study evaluates the accuracy and transferability of these technologies. Using multiple datasets from camera trap environments images were processed with MegaDetector to assess detection accuracy, focusing on filtering empty images and algebraic proof for utilizing per-image assumptions. There was no reliable conclusion for a predictor of best performance using this analysis. Additionally, we analyzed performance using the maximum confidence value within an image. With that, this study highlights the practicality of MegaDetector for blank image sorting, with recall rates on of 95.25%. 85.57%, and 68.76%, respectively, for MD5a, and 95.32%, 85.89%, and 62.53%, respectively, for MDv5b. Despite advancements, this paper emphasizes the continuing need for user-friendly integration into ecological research and continuous validation in diverse environments, recommending selection of technology based on specific project needs and the importance of performance metrics to enhance reliability.
Faculty Advisor/Mentor
Gabriella Dipetto, Michael Pokojovy, Eric Walters
Faculty Advisor/Mentor Email
gdipe001@odu.edu, mpokojovy@odu.edu, ewalters@odu.edu
Faculty Advisor/Mentor Department
Biological Sciences, Mathematics & Statistics
College/School Affiliation
College of Sciences
Student Level Group
Undergraduate
Presentation Type
Poster
Analyzing MegaDetector V5 as a Binary Sorter for Ecological Camera Trap Images
The rapid development of semi-automated workflows within Camera Trap detection has revolutionized the automatic detection and classification of animals. Building upon the foundational work of the MegaDetector object detection model, this study evaluates the accuracy and transferability of these technologies. Using multiple datasets from camera trap environments images were processed with MegaDetector to assess detection accuracy, focusing on filtering empty images and algebraic proof for utilizing per-image assumptions. There was no reliable conclusion for a predictor of best performance using this analysis. Additionally, we analyzed performance using the maximum confidence value within an image. With that, this study highlights the practicality of MegaDetector for blank image sorting, with recall rates on of 95.25%. 85.57%, and 68.76%, respectively, for MD5a, and 95.32%, 85.89%, and 62.53%, respectively, for MDv5b. Despite advancements, this paper emphasizes the continuing need for user-friendly integration into ecological research and continuous validation in diverse environments, recommending selection of technology based on specific project needs and the importance of performance metrics to enhance reliability.