Overview

This repository hosts the source code accompanying the research paper:

Ștefan V., Stark T., Wurm M., Taubenböck H., Knight T.M. (2025). Successes and limitations of pretrained YOLO detectors applied to unseen time-lapse images for automated pollinator monitoring. Preprint (Version 1) available at Research Square, URL: https://www.researchsquare.com/article/rs-6335312/v1 DOI: https://doi.org/10.21203/rs.3.rs-6335312/v1

Pollinator detection in smartphone-captured time-lapse images using NMS-optimized YOLO models that were previously trained on citizen science images (e.g., from iNaturalist, Observation.org).

This GitHub Repository is archived on Zenodo:

How to use this repository

You have two options for accessing this repository: git clone or download it directly.

To clone the repository, ensure that git is installed, then run the following command in your terminal:

git clone https://github.com/valentinitnelav/smartphone-insect-detect.git

Next steps:

• Navigate to ./envs/README.md for instructions on setting up the required Python and R environments.
• Refer to ./data/README.md for instructions on downloading the dataset from Zenodo and details on data processing.
• Consult ./code/README.md for instructions on intermediary processes and model evaluation.
• Check out ./analysis/README.md for instructions on conducting the data analysis.

Data

The image dataset is stored on Zenodo at:

Ştefan, V., Workman, A., Cobain, J. C., Rakosy, D., Wild Stoykova, B., Cyranka, E., Urrego Álvarez, R., & Knight, T. (2025). Dataset of arthropod flower visits captured via smartphone time-lapse photography [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15096610

Check ./data/README.md for more information on how to download the dataset and the data processing steps.

Scripts and notebooks

Scripts and notebooks are organized as follows:

1. ./data/code/: Ground truth data processing. See ./data/README.md;
2. ./code/: intermediary processes to model evaluation. See ./code/README.md;
3. ./analysis/: analysis of the results produced by the optimized detector, along with data descriptors, visualizations, and statistical tests. See ./analysis/README.md.

Reproducibility & Virtual Environments

For Python & R environments, see ./envs/README.md.

This project was developed on Linux using open-source software for the following reasons:

• GPU Compatibility: The GPU workstation and cluster we had access to run on Linux,
• Costs & Open-Source benefits: Linux is free and open-source, enhancing accessibility and transparency. This aligns with our project's commitment to open science and the FAIR principles, allowing others to engage with, review, build upon, and redistribute our work.