Hello, it’s Anne from DeepMind here. Just to clarify - the test set contains 101 labels, which are the 100 turtle IDs from the training set plus an extra ‘new turtle’ label which is used to label any other turtle.

The ‘extra’ dataset can certainly be used in this challenge - we just chose not to use it in the tutorial to keep things as simple as possible. As you’ve noticed, the extra set contains more turtles, and has not been labeled with the image location (top/left/right). 639 of the images in the extra set are additional images of some of the training set turtles.

Here are a couple of initial ideas as to how you could think about using this extra set:

- Use the 639 extra images of the training set turtles to increase the amount of data available for training

- Use all of the extra turtles to try to combat overfitting and improve generalisation

- Use it as a validation set to measure overfitting

I hope that helps!