WildObserve: Camera-trap image review software that learns species labels during annotation
- 1Quantitative Science Consulting, LLC
Abstract
Motion-activated camera traps are an effective tool for monitoring animals, but a single study can produce millions of images, which must be analyzed before they yield data. The burden of analyzing them has led many teams to automate the work with computer vision. These models lose accuracy on images unlike those they were trained on, however, and a general model may not use the categories a project needs. Adapting a model to a project usually requires at least one of the following: many labeled images per species, programming expertise, a graphics processing unit (GPU), or uploading images to a remote server. I describe WildObserve, a desktop application that adapts to each project without any of these requirements. SpeciesNet suggests a species for each animal that MegaDetector finds, and WildObserve keeps SpeciesNet's feature vector for each one. A project model, a logistic regression on those features, then retrains in about a second as the reviewer labels, learning the project's own species and names. From the reviewed records, and optionally the suggestions, WildObserve also fits occupancy and abundance models. Images stay on the user's computer, and every record says which model suggested its species and who confirmed it. WildObserve processed 8.1 images/s on an Apple M1 Pro laptop. On Oregon Critters, a dataset released after SpeciesNet was trained, the project model's accuracy passed that of SpeciesNet with ten labels per species (balanced accuracy 0.783 against 0.751). I also tested occupancy on 63 cameras in Malawi that SpeciesNet had not seen. With 5% of the images reviewed and the project model's suggestions counted for the rest, species' occupancy estimates differed from those of a full review by 0.11 on average. From the reviewed images alone, they differed by 0.41. WildObserve takes a project from images to species records in its own categories and on to occupancy estimates, without writing code, needing a GPU, or sending images elsewhere.
Related
- scPCIP: cross-species annotation of plant cell types with conserved gene programs
- Song Patterns and Stridulatory Movements in Conocephalus Species (Orthoptera: Tettigonioidea: Conocephalidae) of the Eastern Hemisphere: A Review with Some New Data
- Review of the Genus Dolophrades Bates, 1884 (Coleoptera: Cerambycidae: Lamiinae: Morimopsini), with Description of a New Species from China
- Review of Dragon Millipedes (Diplopoda, Polydesmida, Paradoxosomatidae) in the Fauna of Vietnam, with Descriptions of Three New Species
- Physiological mechanisms of drought and high temperature adaptation in Lens species: A review
BibTeX
@misc{tabak2026wildobserve,
title = {{WildObserve: Camera-trap image review software that learns species labels during annotation}},
author = {M. Tabak},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.06.757089},
url = {https://www.biorxiv.org/content/10.64898/2026.10.06.757089v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - WildObserve: Camera-trap image review software that learns species labels during annotation
AU - M. Tabak
PY - 2026
DA - 2026/10/06
PB - bioRxiv
DO - 10.64898/2026.10.06.757089
UR - https://www.biorxiv.org/content/10.64898/2026.10.06.757089v1
AB - Motion-activated camera traps are an effective tool for monitoring animals, but a single study can produce millions of images, which must be analyzed before they yield data. The burden of analyzing them has led many teams to automate the work with computer vision. These models lose accuracy on images unlike those they were trained on, however, and a general model may not use the categories a project needs. Adapting a model to a project usually requires at least one of the following: many labeled images per species, programming expertise, a graphics processing unit (GPU), or uploading images to a remote server. I describe WildObserve, a desktop application that adapts to each project without any of these requirements. SpeciesNet suggests a species for each animal that MegaDetector finds, and WildObserve keeps SpeciesNet's feature vector for each one. A project model, a logistic regression on those features, then retrains in about a second as the reviewer labels, learning the project's own species and names. From the reviewed records, and optionally the suggestions, WildObserve also fits occupancy and abundance models. Images stay on the user's computer, and every record says which model suggested its species and who confirmed it. WildObserve processed 8.1 images/s on an Apple M1 Pro laptop. On Oregon Critters, a dataset released after SpeciesNet was trained, the project model's accuracy passed that of SpeciesNet with ten labels per species (balanced accuracy 0.783 against 0.751). I also tested occupancy on 63 cameras in Malawi that SpeciesNet had not seen. With 5% of the images reviewed and the project model's suggestions counted for the rest, species' occupancy estimates differed from those of a full review by 0.11 on average. From the reviewed images alone, they differed by 0.41. WildObserve takes a project from images to species records in its own categories and on to occupancy estimates, without writing code, needing a GPU, or sending images elsewhere.
ER -