A shiny app for evaluating performance of machine learning models in classifying images of wildlife.

~ 250k photos were identified by human observers as well as a popular pre-trained machine learning algorithm. The study providing the test images closely resembles those used for model training in geographic area, habitat, protocol, and suite of species present.

Actual species and image paths have been anonymized as the data from the test study remains unpublished. The species in the table were made up on the spot as stand-ins for the actual species in the photos, with some attempt made to retain similarites in size and taxonomy.

Because the model was trained on a different dataset the IDs don't correspond to those used in the test data cleanly. IDs in the test data were combined until each ID from the model could be mapped to a single ID in the test data for comparison purposes, but in many cases multiple model IDs correspond to a single human ID. See the "Bird" category in the table for an example. While these photos weren't actually of birds, they represent a related group of species and the model IDs presented as birds here correspond to members of that group.

Some model IDs represent species not present in the test dataset, so mapping them to a correct ID was impossible. These IDs are mostly represented by fish species in the anonymized data, because I was tired of trying to come up with other farm animals.

1. Clone or download the repository from github
2. Open the .Rproj file in RStudio
3. Run packrat::restore() in the console - this step only needs to be performed once, so you can skip it if you open the app again in the future
4. Run shiny::runApp() in the console, or open the app.R file in the editor and click the Run App button