This study explores the effectiveness of YOLOv8 for automatic object detection in behavioral research, particularly in video recordings. YOLOv8, developed by Ultralytics, is noted for its ease of use and high accuracy. The research focuses on surgical tool tracking, a common task in behavioral studies, and examines the conditions required for accurate object detection. Key findings include: 1. **Accuracy with Small Datasets**: YOLOv8 achieves almost perfect object detection even with a small dataset (100 to 350 images), though performance degrades with fewer images. 2. **Background Adaptation**: The detector performs poorly on new backgrounds, but training on a variety of backgrounds significantly improves performance. 3. **Model Size and Version**: Pre-trained models of different sizes (nano, small, medium) show similar performance, with no significant differences in accuracy or computational efficiency. 4. **Training Strategies**: Training a single model on all backgrounds outperforms training separate models for each background, and a subset of diverse backgrounds also yields good results. 5. **Statistical Analysis**: Mixed effects models were used to compare performance across different image set sizes and pre-trained model sizes, showing significant differences in false negatives and intersection over union (IoU) values. The study concludes that YOLOv8 could be a game-changer for behavioral research requiring object annotation in video recordings, provided that appropriate training strategies are employed.This study explores the effectiveness of YOLOv8 for automatic object detection in behavioral research, particularly in video recordings. YOLOv8, developed by Ultralytics, is noted for its ease of use and high accuracy. The research focuses on surgical tool tracking, a common task in behavioral studies, and examines the conditions required for accurate object detection. Key findings include: 1. **Accuracy with Small Datasets**: YOLOv8 achieves almost perfect object detection even with a small dataset (100 to 350 images), though performance degrades with fewer images. 2. **Background Adaptation**: The detector performs poorly on new backgrounds, but training on a variety of backgrounds significantly improves performance. 3. **Model Size and Version**: Pre-trained models of different sizes (nano, small, medium) show similar performance, with no significant differences in accuracy or computational efficiency. 4. **Training Strategies**: Training a single model on all backgrounds outperforms training separate models for each background, and a subset of diverse backgrounds also yields good results. 5. **Statistical Analysis**: Mixed effects models were used to compare performance across different image set sizes and pre-trained model sizes, showing significant differences in false negatives and intersection over union (IoU) values. The study concludes that YOLOv8 could be a game-changer for behavioral research requiring object annotation in video recordings, provided that appropriate training strategies are employed.