Insect Detect platform

Intended use. Generic insect detection model for deployment on Luxonis OAK devices, e.g. OAK-1 used in the Insect Detect camera trap. Works only with uniform, homogeneous backgrounds, e.g. a colored platform.

Architecture and training

Based on YOLOv6 pretrained on the COCO dataset. Trained with luxonis-train on a tiled version [to be published] of the Insect Detect detection dataset.

Inputs and outputs

Expects RGB images with 800x448 px size and NCHW layout as input. Outputs bounding boxes with a single class “insect” and associated confidence scores.

The documented Insect Detect CSV profile follows the v2.0.0 camera-trap metadata writer at revision 24476f4. It saves active tracklets with frame-normalized box corners, display labels, rounded confidence, tracking identifiers and camera metadata. This is source-inspected only: the full deployment probe requires an attached Luxonis OAK device.

Performance

[to be published]

Limitations

Only works on Luxonis OAK devices together with depthai v3. Fails on complex backgrounds (e.g. vegetation) and requires camera to object distance of < 30 cm for reliable detection.

How to obtain the weights

Weights in Luxonis NN Archive format published as release asset. Other formats including PyTorch and ONNX will be published soon.

Citation

@software{sittinger_2026_19372336,
  author       = {Sittinger, Maximilian},
  title        = {Insect Detect - Software for automated insect monitoring with a DIY camera trap system},
  month        = apr,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v2.0.0},
  doi          = {10.5281/zenodo.19372336},
  url          = {https://doi.org/10.5281/zenodo.19372336},
}