Hybrid marmosets are being released in areas causing damage to the native marmosets population in the local ecosystem. Identification of these hybrids is vital to protecting the local endangered species of marmosets. We aim to train state of the art image classification models to classify hybrid marmosets from parent marmoset species. The project involves image annotation, color standardization, model training, and mobile app creation.
First install required packages:
pip install -r requirements.txt
dectectron2 must be installed separately after:
pip install 'git+https://github.com/facebookresearch/detectron2.git'
Note: If running on Windows, may need to install Windows C++ build tools from https://visualstudio.microsoft.com/visual-cpp-build-tools/. This is needed to install detectron2.
The images used to train this model are compiled in the hybrid_photos dataset on Hugging Face. To access the data, run
python download.py --dest_path <filepath> --repo_id <repoID>
Then, to generate the model, run
python train.py
This will run the appropriate scripts to process the data for training and train the model.
Testing can be done with run_tests.py, though some paths are hard-coded.
The app takes in an image of a marmoset and runs it through our classifier to determine its species (the confidence of the prediction is also provided).
Additionally, the app displays the uploaded image alongside a sample image of the predicted species. It informs the user of the predicted species' natural habitat or that it is "exotic", meaning that it is a hybrid, and therefore has no natural habitat.
There is an actively running instance of the marmoset classifier app available here. The user simply uploads a JPG or PNG image, and the results will be displayed below. Note that all information is session-specific, so refreshing the page will clear the app cache.
If you want to run your own instance of the app, follow the instructions below.
First install required packages:
pip install -r app/requirements.txt
Note: The requirements for the app are not the same as the requirements for the model.
The sample images come from our training set, and are compiled in the hybrid_photos dataset on Hugging Face.