Animal detection with Faster R-CNN: Transfer Learning and Domain Adaptation

A Multi-domain generalizable animal detection model: exploration of potential solutions to improve the SOA.

Table of content

• Requirements
• Getting started
• Examples of experiments
• Authors

Requirements

• Python 3.7.11
• An ipython notebook editor and runner (Jupyter Notebook, VSCode with extension, etc...)
• Having read our report

Getting started

1. Download the Git repo using:

   git clone https://github.com/Hazot/Animal-Detection-and-Generalization.git
   

   or download it manually.

2. Download the dataset (Benchmark images: 6GB; Metadata files: 3MB) under the header CCT20 Benchmark subset: https://lila.science/datasets/caltech-camera-traps. Extract the dataset directly into the project folder. Make sure to have the following folders:

• eccv_18_all_images_sm
• eccv_18_annotation_files

4. Load the virtual environment librairies using:

   python -m pip install -r requirements.txt
   

5. Open the main_notebook.ipynb and simply execute the cells sequentially.

Examples of experiments

1. To reproduce the test results of a RPN+ROI model, simply execute the cells until the interactive part. Make sure that you use the right parameters:

   data_augmentation_mode = 'none'
   model_depth = 3
   

   In the cell below:

   lightweight_mode = 1
   

   This will use smaller datasets and only 5 epochs. Exercute every cells until right before the optional part (Make Predictions with a model). The results if this training will not be useful at all. To get better results, you need a lot of time (4-5 hours) and disable the ligthweight_mode by doing:

   lightweight_mode = 0
   

2. To use domain adaptation on this model, after training a model go to the Domain Adaptation heading and run every cell below until the very last.

Authors

• Abdiel Fernandez ([email protected])
• Rose Guay Hottin ([email protected])
• Kevin Lessard ([email protected])
• Santino Nanini ([email protected])