Deep Metric Learning

In the image matching task, a model receives a pair of images and determines if they are similar or not. We use Deep Distance Metric Learning to extract a lower dimensional representation of images and measure their similarity. The goal is to cluster similar images in a K-dimensional space. Two main sub-problems are identifying an objective function and designing a CNN feature extractor. We evaluate different approaches and CNN architectures in this method. Our training procedure and techniques to speed up convergence are also presented and evaluated.

Janek Zimoch
Janek Zimoch
Data Scientist

Related