Robustness tests: non-overlapping test/train datasets

[samcunliffe]

We'd like to build a more generalisable model so that e.g. the model would work on fresh data from a similar beach but unseen videos.

Definition of done

• The dataloader is modified so that the training and evaluation datasets are non-overlapping (frames from the same stereo pair).
• The user can turn this on or off somehow.

Possible implementation:

• Define the test and training datasets with indices (belonging to the same pair of cameras.