Deep metric learning(DML) aims to find an feature embedding space for the images such that images of the same category are closer to each other than images of any other category.
Given only labeled data from the source domain(synthetic data), the goal is to learn features such that they transfer to the target domain(real images), which has no labels. The aim is to align the feature distribution of source domain and target domain.
[To be updated] The key questions for image classification deal with selecting an architecture, an appropriate loss function, an optimizer, and how to solve the challenge unbalanced classes. After getting the proof of concept(PoC) right, the next step is to successfully train a model?