beetee17/few_shot_learning

DSTA Internship Project Apr - Jun 2021

few_shot_learning's Issues

200421: preprocessing of military dataset

The total size of the dataset is ~1500 images. The advantage of siamese network is that it is trained on pairs of images rather than a single image. This means there are 1500 choose 2 possible samples that can be extracted from the dataset.

However, doing so would mean that a large majority of our samples would be negative samples. Unsure if I should proceed to modify make_pairs function to include many more samples, even if it might mean non-uniformity of samples.

Current implementation of the function yields equal distribution between positive and negative samples, but only produces 2*1500 total samples