Bjerge et al. (2023)

YOLOv5 model trained on flower pollinators.

Intended use. Detecting pollinators visiting flowers in field imagery. It is built for the cluttered-background case where the insect is a small fraction of the frame and the flowers are visually dominant.

Architecture and training

• Architecture: YOLOv5
• Task: Object Detection (bounding boxes)
• Trained on: flower pollinator imagery, see the source publication for the dataset composition
• Taxonomic coverage: pollinators, see the source publication for the exact class list
• Framework: ultralytics YOLOv5

Inputs and outputs

• Input size: 1280x1280, the authors report this as the best-performing configuration
• Channel order: RGB
• Outputs: bounding boxes
• Input modality: image only (discriminative model)

Performance

See the source publication.

Limitations

License and rights

• License: CC-BY-4.0
• Commercial use: allowed
• Restrictions: Attribution is required by the license. Cite the source publication listed under Citation below.
• Upstream source: https://zenodo.org/records/7395752

Citation

@article{bjerge2023pollinator,
  title   = {Detection of flower-visiting insects with deep learning},
  author  = {Bjerge, Kim et al.},
  journal = {PLOS Sustainability and Transformation},
  year    = {2023},
  doi     = {10.1371/journal.pstr.0000051}
}

How to obtain the weights

From Zenodo record 7395752. No authentication is required. The record contains several files, the weights are in the YOLOv5models.zip file.

Output format probe

The published 1280s6 checkpoint was tested with the author’s insectsFlowers CSV exporter. It produced headerless CSV rows with pixel corners, one-based class IDs and integer percentage confidence, alongside YOLO text. That repository describes a later study; its historical pairing with this 2023 model is unconfirmed, so the advertised format remains unassigned. Probe date: 2026-09-30.