| Summary: | Data are available to download at: https://arcticdata.io/data/10.18739/A2J38KJ9R/ The image subset and data provided here are from 20 sampling locations from an ongoing camera trap project (2019-2023) in Arctic Alaska, United States. We used this image set to evaluate a state-of-the-art computer vision model developed by Microsoft AI for Earth named MegaDetector for detecting wildlife in camera trap images. To better understand the utility of automated data processing tools for wildlife camera trap data, we compared image labels (i.e., images with/without animals) assigned manually by human reviewers with those determined using MegaDetector. We found MegaDetector reliably determined the presence or absence of wildlife in camera trap images generated by motion detection camera settings (e.g., ≥95% accuracy), however, performance was substantially poorer for images collected with time-lapse camera settings (e.g., ≤62% accuracy) as compared to human review. Our results highlight the importance of making mindful decisions about how to collect and process camera trap images and the impact those decisions have on the information generated to inform study outcomes. This human-labeled dataset may also be used to evaluate and improve future development of computer vision models for automated image analysis.
|
|---|