Using TensorFlow to recognize animals in infrared camera videos

Dennis Sosnoski is a deep learning engineer working with The Cacophony
Project (https://cacophony.org.nz/) on developing TensorFlow models to
automatically classify sounds and images. An important part of the work being done by Cacophony is monitoring and trapping of introduced predator
species, and as part of this they're sourcing infrared video cameras to
individuals and organizations to observe predator activity in the
environment.

Previously the recordings from these cameras had to be viewed and tagged by a human for accurate identification of animals caught on-camera, but new models based on deep learning are allowing Cacophony to switch to automated tagging of the videos. This makes it practical to deploy the cameras to monitor more territory at lower cost., bringing us all a big step closer to quantifying the predator problem in Aotearoa and finding the best solutions.

In this talk Dennis will dig into the details of the models being used for recognizing animals in infrared video recordings. He'll discuss the main models and data representations attempted, including InceptionResNetV2, ResNet variations, and EfficientNet. He'll also show how the data is prepared and presented to TensorFlow for training the models, and how different variations have performed to date.

This meeting is an in-depth follow-on to the Tuesday meeting https://www.meetup.com/machine-learning-data-science-WLG/events/279136948/?isFirstPublish=true The Tuesday meeting provides context and summary information, this one gets into the details. Links to the source code for the main models will be provided prior to the meeting to give everyone a chance to review and make suggestions or raise questions.