Mothbot Detect (MBD)

Finds every insect on a Mothbox light-trap photo and returns one oriented box per individual.

Intended use. The detection step of Mothbot, the processing software for the open-hardware Mothbox light trap. It is meant for whole-night series of high-resolution photos of an illuminated sheet, which can carry thousands of insects per frame. The detector does not name anything. Mothbot passes each crop to BioCLIP 2 (zero-shot, see the BioCLIP 2 entry) and filters the result to a GBIF species list for the region.

Architecture and training

• Architecture: Ultralytics YOLO26 with oriented bounding boxes (OBB), one class (creature).
• Checkpoints: MBD-1-1 (September 2026, default) and MBD-1-0 (July 2026). An older MBD-0-2.pt without a config file is also in the repository.
• Training data: the README describes the interim v0.1 model as trained on 7,000 insects at imgsz 1600, from Mothbox deployments in Panama, Mexico, Croatia, Germany, Poland, Hawaii, North Carolina, the Netherlands, Indonesia, Canada and Seattle. The original 2025 YOLO11-OBB model was trained on 4,500 insects, mostly from Panama. The training data for MBD-1-0 and MBD-1-1 is not documented. The display name “15kbugs” hints at about 15,000 insects, but this is not stated. The training set is not published.

The rest of the Mothbot pipeline uses off-the-shelf models, which are not part of this entry:

• BioCLIP 2 (imageomics/bioclip-2 via pybioclip), with order rank as the default.
• DINOv2 ViT-S/14 embeddings, clustered with HDBSCAN.
• BiRefNet (via rembg), which segments crops for a pixel-mass size estimate.

Inputs and outputs

• Input: RGB light-trap photos, run at imgsz = 1600.
• Shipped YAML settings: the YAMLs set conf = 0.1 and max_det = 5000 (MBD-1-1) or 500 (MBD-1-0).
• Output: one JSON per image in X-AnyLabeling/LabelMe style. Each detection is a shape with label: "creature", shape_type: "rotation" and four corner points. Cropped patches are saved next to it.
• Later pipeline steps add the taxon, the ID confidence and the cluster ID to the same JSON, and finally write a Darwin-Core-like CSV.
• Models: .pt files run through ultralytics.YOLO. ONNX export is referenced in the YAMLs, but the ONNX files are not in the repository and ONNX selection is disabled in the UI.

The documented Mothbot detection JSON profile follows the PyTorch OBB writer at revision 7514bee. A CPU probe of MBD-1-1 and the original _save_result produced 15 shapes on a specimen fixture and an empty shape list on a blank control. The source invocation uses imgsz=1600 and max_det=10000, leaving confidence/IoU to the model library; YAML training settings are not explicitly forwarded here. The probe covers inference and JSON serialization, excluding thumbnail generation, optional blur enrichment and later identification/tracking stages.

Performance

Not reported in the source publication. There is no mAP, precision or recall, and no evaluation set, in the repository or on the documentation site.

Limitations

• Class-agnostic. Everything on the sheet is a creature: moths, beetles, flies, spiders, debris that looks like an insect. Any taxonomy comes from the downstream BioCLIP 2 step. Its docs say that step is reliable mainly at order or family level.
• Domain. Trained on Mothbox sheet photos at about 1600 px. Expect worse results on other backgrounds, other trap geometries, or much lower resolution.
• Undocumented checkpoints. The README still describes the 2025 YOLO11 model and the v0.1 interim model. It says nothing about the MBD-1-x checkpoints in this entry. For MBD-1-1, the model size is ambiguous: the YAML says s, the commit says M.
• Performance. No evaluation has been published.

How to obtain the weights

The .pt files are in the trained_models/ folder of the GitHub repository. The asset links above are pinned to commit 7514bee8ace5c9cb859f3566ed3088a7824b692c. They are also bundled in the desktop-app releases. No authentication is required.

License: the Mothbot_Process repository has no license file. Following the convention used for Stark et al. (2023), this entry records AGPL-3.0, the license of the Ultralytics YOLO code the weights were trained with. If the authors add a license file, that file is authoritative.

Citation

No paper or citation file exists. Cite the software:

@software{mothbot_process_2026,
  author = {{Digital Naturalism Laboratories}},
  title  = {Mothbot\_Process: moth detection workflows and desktop tooling},
  year   = {2026},
  url    = {https://github.com/Digital-Naturalism-Laboratories/Mothbot_Process}
}