This goal of this project is to provide researchers at Department of Natural Resources and Renewables (NRR) in Nova Scotia with an easy tool to process large camera trap datasets. The tool includes an object detection deep learning model as well as a front-end tool to make is easily accessible without the need of technical knowledge. The project is being developed as a part of Computer Science honours program at Memorial University of Newfoundland.
I developed the model using a popular machine learning python library called Pytorch. PyTorch is a machine learning library based on the Torch library, used for applications such as computer vision and natural language processing. The model uses a popular object detection model called FasterRCNN and with a pretrained backbone from ResNET. I am basically retraining this model to detect the local species in Nova Scotia.
The model is being trained periodically over a data set provided by NRR. The dataset includes about 100k images of 23 different species in Nova Scotia.
The images needed to be pre-processed by generating bounding boxes for the object detection task. I used MegaDetector to generate the bounding boxes. MegaDetector is an object detection model made by Microsoft. It classifies the detected objects into humans, animals, and vehicles.
The first tested image 🙂