Camera traps are a viable tool for ecologists to record and monitor the behavior of wildlife. The detection of presence and species of animals in the acquired images are two examples of semantics that are of central interest to researchers. Traditionally, labeling each trap image with such information has been carried out manually, which is a tedious and expensive task to do. In this work, we aim at solving this problem automatically by means of state-of-the-art machine learning tools.
Automated vision tasks, such as the classification of objects in still images, have seen tremendous progress in the respective research fields of Computer Vision and Machine Learning. This can be traced back to the introduction of Convolutional Neural Networks (CNNs), a family of powerful and versatile machine learning algorithms. They work by being trained on extremely large, annotated datasets that contain natural images of everyday objects.
The theoretical concept of CNNs enables us to also use them for classifying animals and their species in camera trap images. Due to the lack of training labels, this is a challenging task.
In this project, the student will attempt at training a CNN to detect the presence of animals and classify their species in camera trap images. To overcome the label shortage issue, they will investigate one of the most recent research branches of Selfsupervised Learning, where a model trains itself on a task that does not require any labels. Upon completion of this first stage, the model is pre-trained and can be adapted to the classification target, overall yielding superior performance without the need of large-scale training datasets.
- Familiarize and successfully setup a CNN-based animal classifier for camera trap images
- Investigate different self-supervised learning techniques
- Provide a machine-estimated set of animal presence and species, along with an accuracy score
- Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
- Wang, Xiaolong, Kaiming He, and Abhinav Gupta. "Transitive Invariance for Selfsupervised Visual Representation Learning." arXiv preprint arXiv:1708.02901 (2017).
- Gomez, Alexander, Augusto Salazar, and Francisco Vargas. "Towards automatic wild animal monitoring: identification of animal species in camera-trap images using very deep convolutional neural networks." arXiv preprint arXiv:1603.06169 (2016).
- Programming skills in MATLAB or Python (or high motivation for learning);
- some background in statistics and/or machine learning is an asset.
Theme: Modelling & visualisation