Automated location invariant animal detection in camera trap images using publicly available data sources

Area of Science:

• Ecology
• Computer Science
• Artificial Intelligence

Background:

• Camera trap data analysis is time-consuming and resource-intensive.
• Current automated image classification software lacks location invariance, hindering large-scale deployment.
• Retraining deep learning models for new sites requires significant annotation effort.

Purpose of the Study:

• To develop a method for creating location-invariant camera trap object detectors.
• To evaluate the effectiveness of publicly available image datasets for training.
• To optimize models for high-accuracy, domain-specific ecological applications.

Main Methods:

• Collected and annotated datasets from Flickr and iNaturalist (FiN) for striped hyena, rhinoceros, and pigs.
• Trained object detection models using FiN datasets and compared them to models trained on existing camera trap datasets.
• Optimized model robustness by infusing small subsets of camera trap images into the training data.

Main Results:

• FiN-trained models achieved significantly higher mean Average Precision (mAP) (82.33%-88.59%) compared to models trained solely on camera trap data (38.5%-66.74%).
• Infusing camera trap images further improved mAP by 1.78%-32.08%.
• The developed method enables robust, out-of-the-box object detection software for ecological applications.

Conclusions:

• Publicly available image datasets can be effectively used to train robust, location-invariant deep learning models for camera trap image processing.
• Infusing a small percentage (5%-10%) of site-specific camera trap images enhances model performance.
• This approach facilitates the large-scale deployment of AI technologies in ecological management and research.