Future Work

Given the time and hardware constraints, we didn’t get to try out other ideas that could potentially give us a better-performing model. Below are the list of adjustment to our current model that we think would improve the accuracy of our model on a more general test dataset.

• Make a composite train set from augmenting images of the original dataset and try RESNET50 or maybe even RESNET101 (on colab pro).
• Try averaging the weights of two models, one trained on the base train set and one trained on a very augmented data set.
• Try VGG16 model. Some forums online suggested they saw less overfitting with this model.

And, if all things fail…

• Hire a team of 50+ professional birdwatchers to classify the birds for us.