POLLINATOR — Serra-Marin et al. (2025)

Intended use. Find flower-visiting insects in field images and video frames for subsequent human review. Flowers and vegetation make this a detector for heterogeneous backgrounds.

Architecture and training

The repository describes fine-tuning pretrained YOLOv5m weights. Its dataset configuration defines one class, pollinator. The MERGED-ALL v4 export documents 16,389 images, orientation correction, stretching to 1024 × 1024 pixels, and augmentation with flips and rotations. This describes the repository’s export; its exact correspondence to the downloaded checkpoint is unverified.

Static inspection on 2026-09-29 of the checkpoint now published as pollinator_serra_marin_2025_best.pt confirms weights = yolov5m.pt, one class (0: pollinator), and architecture depth/width multipliers of 0.67/0.75. Its training options record imgsz = 1024, epochs = 150, and SGD. These are configured values, not proof that all 150 epochs completed: the saved epoch is -1, and best_fitness and optimizer state are absent (None).

The checkpoint records dataset path ./NEW_DATA_TOT/data.yaml, run directory runs/train/exp4, timestamp 2025-06-19T22:26:14.768546 (no timezone), and Ultralytics YOLOv5 commit 17c500461d7b14a24133d91bc6437af62914074c. It does not embed that dataset’s image list or split membership.

Inputs and outputs

• Input: RGB images; 1024 × 1024 is the documented training size.
• Output: bounding boxes, confidence scores and the pollinator class.
• Video processing: the supplied script converts OpenCV frames to RGB and saves frames into with_object and without_object folders. It calls model(img_rgb) without explicitly setting inference size; the training size is not enforced by that script.

Performance

Table 2 of the paper reports five-fold averages for these combined-data experiments:

Neither the repository nor the inspected checkpoint establishes which table row or fold corresponds to pollinator_serra_marin_2025_best.pt; these are study results, not verified checkpoint-specific scores. No inference or benchmark was run during inspection.

Limitations

Small insects, leaves mistaken for insects, and unseen plant backgrounds cause errors. Species identification and confirmation of flower contact require human review. A detection alone does not establish a pollination interaction.

How to obtain the weights

The weights can be obtained from either location:

• Google Drive: download the checkpoint from the authors' weights folder, linked in the original repository’s download instructions as best.pt. This link is not version-pinned.
• Hugging Face: download pollinator_serra_marin_2025_best.pt. This download is pinned to commit 900afc5b and requires no authentication.

The renamed file matches the previously inspected checkpoint: it contains 42,336,660 bytes and has SHA-256 9b62c544b5dc31122d93db8ea9363aa4012a1b7636e57de740fa6ae6a2cc600e. Use these values to check whether a later download matches this checkpoint. The Zenodo v1.0 release archives the code and download pointers, not the checkpoint itself.

License and rights

Zenodo declares CC-BY-4.0 for the software archive, credited to Arancha Lana. The dataset configuration also declares CC BY 4.0. The Hugging Face model card declares AGPLv3 (agpl-3.0) for the model repository containing pollinator_serra_marin_2025_best.pt.

Citation

@article{serramarin2025pollinator,
  title   = {Comparative assessment of automated and manual monitoring in comprehensive plant–pollinator communities},
  author  = {Serra-Marin, Pau Enric and Solé-Ribalta, Albert and Lana, Arancha and Borge-Holthoefer, Javier and Hervías-Parejo, Sandra and Traveset, Anna},
  journal = {Methods in Ecology and Evolution},
  volume  = {16},
  number  = {12},
  pages   = {2960--2978},
  year    = {2025},
  doi     = {10.1111/2041-210X.70165}
}

Software: Arancha Lana (2025). aranchalana/POLLINATOR: First stable release of POLLINATOR code (v1.0). Zenodo.

Output format probe

The author video script was run with the published checkpoint. It saves JPEG frames into with_object and without_object folders and exports no machine-readable boxes. The tested frames exercised without_object only; the annotated with_object branch is documented by the source.

Probe date: 2026-09-30. Reproducible setup and evidence are in src/probe.