Overview

Deep Metric Learning (DML) research is plagued by inconsistency — papers with different training protocols, architectures, and parameters make it impossible to compare techniques. To address this:

1. We benchmarked models from the top X papers using a standardized setup.
2. We found a strong correlation between embedding space density and compression and generalized model performance.
3. We propose a new training regularization to reliably boost model performance.

Quick Links

• Code on GitHub →
• Paper on arXiv →

Image Similarity Tasks

DML is valuable for image clustering and image retrieval. Models learn an embedding space that accurately captures similarities between images using a defined distance metric. For example, this can be used to find similar images of cells under a microscope, making it easier for physicians to diagnose diseases (Yang et al).

Image from https://github.com/easonyang1996/DML_HistoImgRetrieval

Comparing metrics and performance

This set of plots shows ... [todo]

Each line represents a grouped set of experiments, where each experiment has a different random seed. [todo ?]

Further findings

Here, you can see our next set of experiments where ... [todo]

Further research directions

We recommend using the new ...