This set of notebooks is designed to introduce ocean scientists to computer vision tools for working with image and image-like data.
Participants will learn basic concepts in computer vision, from manipulating individual images to semantic classification with convolutional neural networks. The material focuses on plankton images collected with the Scripps Plankton Camera System and the ZooScan. Later lessons use the LILA BC Ohio Small Animals dataset to introduce the commonly used COCO data format and advanced techniques. Each module treats different, related topics but can be used independent of one another.
The data sets have all been preloaded on the NOC Data Science Platform (DSP). There are some hard coded file paths that assume you are running the examples on the DSP.
$ conda run -p /groups/cv-workshop/cv-workshop-env python -m ipykernel install --user --name cv-workshop
After about one minute a new cv-workshop icon will appear on your launcher and in the list of available kernels.
Only do this if you aren't using the Data Science Platform or if the above fails. It will take a few minutes and use several gigabytes of storage (DSP users only get a 20GB quota by default).
This set of notebooks uses the compute environment described in environment.yaml. To create the environment on the DSP, open a terminal from the Launcher and run:
$ conda env create -n cv-workshop -f environment.yaml
This creates a new named environment from the configuration file.
To make the environment visible to your Jupyter notebooks run:
$ python -m ipykernel install --user --name cv-workshop
Within a few minutes, the environment should appear in your Jupyter launcher.
Note: The original material was developed for the Partnership for Observing the Global Ocean's 2019 Workshop on Machine Learning and Artifical Intelligence in Biological Observations by Eric Orenstein and Simon-Martin Schroder (University of Kiel).
Authors: Eric Orenstein (National Oceanography Centre), Adam Ward (ORCID) (National Oceanography Centre)
Maintainers: Adam Ward