Mothbox Training Dataset 1-1

MOTHBOX TRAINING DATASET 1.1

An annotated image dataset of nocturnal insects photographed at light
attractors by Mothboxes. It was used to train the Mothbox insect detection
model, and we are sharing it so others can reuse it for their own detection
and classification work.

Only one class being ID'ed- "creature."

Our general rubric for annotating was to label everything in a source image that we were sure was a living creature of some kind in the Kingdom Animalia (vs background or dirt or frass or plants). Thus, this includes extremely small insects, insects blurred due to their motion across a rolling shutter, insects blurred due to camera artifacts or weather, but we could still infer that they were animals of some sort.

Annotated by Kit Quitmeyer and Andy Quitmeyer
Digital Naturalism Laboratories

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

ABOUT THE DATA

Collection: Images were captured with Mothboxes of various models and
community remixes, deployed in 14 countries between 2024 and 2026.

Annotation: Insects were annotated with oriented bounding boxes (OBB) in
X-AnyLabeling.

Format: The dataset is shared in YOLO OBB format: JPEG images, one .txt label
file per image (class x1 y1 x2 y2 x3 y3 x4 y4, normalized coordinates), and a
data.yaml file describing the class and the train/val/test splits.

DATASET STATISTICS

Images:        1,695 full-resolution JPEGs (about 26 GB)
Annotations:   14,549 oriented bounding boxes
Classes:       1 ("creature", meaning any animal (generally arthropod) at the attractor)

Split     Images   Annotations   Background images (no insects)
train      1,356        11,916      256
val          169           939       21
test         170         1,694       25

Image resolutions:
  9248 x 6944 px (64 MP): 1,187 images
  9000 x 6000 px (54 MP):   508 images

Annotations per image: mean 8.6, median 2, max 1,781
Box size (square root of box area): median about 111 px, middle 90% of
boxes between 41 px and 552 px, full range 14 px to 2,964 px

The 302 background images contain no insects. They are included on purpose
as negative examples, to reduce false positives on empty sheets, debris,
and glare.

DATA COLLECTED BY

Andrew Quitmeyer: Panama, Ecuador, Peru, USA, Italy, Germany, Denmark, Croatia, Indonesia
Kit Quitmeyer: Panama, Ecuador, Peru, USA, Italy, Germany, Denmark, Croatia
Hubert Szczygiel: Ireland, Panama, Poland, USA (Hawaii, Alaska)
Bri Johns: USA (NY, NC), Panama, Indonesia

Wira Artawan: Indonesia

Wilan: Indonesia

Edgar Bernal: Indonesia
Logan Williams: USA (NY)
Chris Gould: USA (WA)
Vittorio Saggiomo: The Netherlands
Serina Tourangeau: Canada (Burlington, Ontario; summer 2026; red oak and
  hickory forest)
Richard Feldman: Canada (Carden Alvar Provincial Park, Ontario; May-July 2025)
  Project lead / principal investigator: Richard Feldman, Research Scientist,
  Wildlife Research and Monitoring Section, Ontario Ministry of Natural
  Resources
  Mothbox builds and operation: Noah Opper, Ontario Ministry of Natural
  Resources
Juan Hurtado: Argentina (Tymbachui, Misiones)
Sil A. Belliard: Argentina (Tymbachui, Misiones)
Jessie Tejara-Fabian: Panama
Jorge Medina: Panama
Jeff Dietrich: Panama
Mulget Amaru: Panama
Sol Parra Santos: Panama

FUNDING AND SUPPORT

Research funding was provided primarily by:
  Wildlabs.net
  ARM
  The American Museum of Natural History

Additional support from:
  Pro-Eco Azuero
  Ponterra
  University of Washington (project storage space)
  IVADO Institute (sponsored a collaboration with Antenna software on
    insect image processing)