PROJECT TITLE :

Sharing Matters for Generalization in Deep Metric Learning

ABSTRACT:

The foundation for a wide variety of vision tasks is the acquisition of an understanding of the similarities between different images. The dominant paradigm is known as discriminative metric learning, and its goal is to find an embedding that differentiates between various training classes. However, the most difficult part is learning a metric that not only generalizes from training to novel test samples, but also generalizes to test samples that are related to the training samples. Additionally, it ought to transfer to a variety of object classes. Therefore, what complementary information does the discriminative paradigm not take into account? We need to find characteristics that, in addition to separating different classes, are also likely to occur in novel categories. We can tell that this is the case if the characteristics are shared across different training classes. In this body of work, we investigate ways to learn such characteristics that do not require the use of additional annotations or training data. Our method, which we have recast as an innovative triplet sampling strategy, can be quickly applied on top of more recent ranking loss frameworks thanks to the formulation we have given it. Experiments show that our method significantly improves performance in deep metric learning, leading to new results that are state-of-the-art on a variety of standard benchmark datasets. This improvement is independent of the network architecture that is underlying the learning and the specific ranking loss that is being used.