Version 2 2025-01-22, 21:39Version 2 2025-01-22, 21:39
Version 1 2024-02-20, 20:32Version 1 2024-02-20, 20:32
model
posted on 2025-01-22, 21:39authored byAmira BurnsAmira Burns, Ryan Miller, Hailey Wilmer, Michael Tabak, Daniel Falbel, Tess Hamzeh, Ryan K. Brook, John A. Goolsby, Lisa D. Zoromski, Raoul Boughton, Nathan Snow, Kurt VerCauteren
CameraTrapDetectoR is an R package that uses deep learning computer vision models to automatically detect, count, and classify common North American domestic and wild species in camera trap images. Data for all versions of the taxonomic species model are located in this dataset. This data is automatically downloaded, extracted, and deployed in the tool's deploy_model function. Additional information about the R package and the training data can be found in the package's Github repository: https://github.com/CameraTrapDetectoR/CameraTrapDetectoR
This research used resources provided by the SCINet project and the AI Center of Excellence of the USDA Agricultural Research Service, ARS project number 0500-00093-001-00-D.
List of Resources:
species_v3.zip [added Jan 2025] is a folder containing the model weights, architecture, and class label dictionary for the third version of the species model. The model is a YOLOv8 architecture trained on the ARS SCINet Atlas cluster; it identifies and counts 64 North American species or species groups in camera trap images, including humans vehicles and empty images. The training set contains 134,845 unique images across the 64 classes collected from 32 sites.
species_v3_cl.zip [added Jan 2025] contains is a folder containing the all information to deploy the species v3 model via Python script from the command line. Full instructions for set up and use may be found at https://github.com/CameraTrapDetectoR/model_training
species_v2.zip is a folder containing the model weights, model architecture, and class label dictionary for the second version of the species model. The model architecture is a FasterRCNN object detection model with a ResNet50 backbone, trained on the ARS SCINet Atlas cluster. This model identifies and counts 78 North American species in camera trap images, including humans vehicles and a background class. The training dataset contains 169,352 unique images, with an average of 2199 images per class excluding background class. The (min, max) range of images count per class is (107, 7027); this class imbalance was addressed with a suite of data augmentations and weighted random sampling. Images were acquired from a total of 26 databases across North America.
species_v2_cl.zip is a folder containing the all information to deploy the species v2 model via Python script from the command line. Full instructions for set up and use may be found at https://github.com/CameraTrapDetectoR/model_training
species_v1.zip is a folder containing the model weights, model architecture, and class label dictionary for the first version of the species model. The model architecture is a FasterRCNN object detection model with a ResNet50 backbone.
Funding
USDA-ARS: 0500-00093-001-00-D
USDA-APHIS: Center for Epidemiology and Animal Health
The dataset supports the R package CameraTrapDetectoR with model architecture, model weights, and class label dictionary for the taxonomic species model version 1, launched January 2022, version 2, launched May 2023, and version 3, launched January 2025.