YOLOBee: Tracking and Analyzing Bee Behavior Using YOLO Models
Description
- Bees are crucial for biodiversity through their role in pollination, which supports plant reproduction and global food production.
- The YOLOBee project aims to study bee trajectories and preferences for different flowers using YOLOv5 and YOLOv8 models to improve bee detection and tracking.
- Experimental setup involved 20 3D-printed flowers with controlled parameters.
- Training involved multiple YOLOv5 models (s, m, x) and comparisons with YOLOv8 models, using diverse datasets and augmentations.
Folders and Contents
- Training Folders (00-06):
- Progressive improvements and comparisons of YOLOv5 models with varying datasets and augmentation techniques.
- Results include training, validation, and detection outputs.
- YOLOv8 Folder:
- Comparison of YOLOv8 models with previous YOLOv5 results.
- Notebook Folder:
- Google Colab notebooks for augmentation, training, validation, and detection.
- Ready-to-use examples for bee detection.
- Notes Folder:
- Documentation of various project tasks.
- Videos Folder:
- Videos used for detection steps.
- Dependencies:
- Details of Python and package versions used in the project.
- Links to Related Projects:
- References to external projects and tools used (YOLOv5, Ultralytics, SciCount, SciAugment, Watermark).
Workflows
Data Preparation
- Dataset creation and annotation using software like makesense.ai.
- Augmentation to generate additional training images.
Training
- Steps to train the neural network, including parameter explanations and GPU usage.
- Comparison of different YOLOv5 and YOLOv8 models.
Detection
- Process for detecting bees in videos or frames.
- Output includes text files with detected objects and their coordinates.
Results and Comparisons
- Detailed analysis of training results, including confusion matrices and mAP scores.
- Comparisons between YOLOv5 and YOLOv8 models.
- Tracking bee movements using Trackmate software, visualizing trajectories and speeds.
Contributors
We would like to thank Jakub Štenc for pitching the idea for this project and providing us with the video datasets!
Acknowledgements
Computational resources and consultations were provided by the Vinicna Microscopy Core Facility co-financed by the Czech-BioImaging large RI project LM2023050. Additional computational resources were provided by the e-INFRA CZ project (ID:90254), supported by the Ministry of Education, Youth and Sports of the Czech Republic.