Automatic object detection for behavioural research using YOLOv8.

Area of Science:

• Computer Vision
• Behavioral Science
• Machine Learning

Background:

• Observational studies of human behavior frequently require manual object annotation in video recordings.
• Advancements in automatic object detection, particularly YOLOv8 (You Only Look Once version 8), have simplified this process.
• YOLOv8 is recognized for its ease of use and efficiency in object detection tasks.

Purpose of the Study:

• To investigate the specific conditions necessary for achieving accurate object detection using the YOLOv8 model.
• To evaluate the performance of YOLOv8 under varying training dataset sizes and background conditions.
• To determine the potential impact of YOLOv8 on the efficiency and accuracy of object annotation in behavioral research.

Main Methods:

• Utilizing the YOLOv8 object detection model for analysis of video recordings.
• Training YOLOv8 on datasets of varying sizes, ranging from 100 to 350 images.
• Assessing model performance with objects presented against consistent and diverse backgrounds.

Main Results:

• YOLOv8 demonstrated nearly perfect object detection accuracy when trained on small datasets.
• The model exhibited limitations in generalizing object detection to new backgrounds when trained on limited background variety.
• Restoring excellent object detection performance was achieved by incorporating a wider range of backgrounds during model training.

Conclusions:

• YOLOv8 shows significant promise for behavioral research requiring object annotation in video data.
• Dataset diversity, particularly background variation, is crucial for robust YOLOv8 performance.
• The ease of use and high accuracy make YOLOv8 a potentially transformative tool for researchers.