Attend to Context for Refining Embeddings in Deep Metric Learning

Dmytro Kotovenko, Pingchuan Ma, Marcel Fischer, Björn Ommer

Keywords: deep metric learning, supervised representation learning, survey

Abstract: The primary objective of deep metric learning (DML) is to find an effective embedding function that can map an image to a vector in the latent space. The quality of this embedding is typically evaluated by ensuring that similar images are placed close to each other. However, the evaluation step, which involves finding the sample and its neighbors and determining which neighbors share the same label, is often overlooked in the current literature on DML, where most of the focus is placed on training the embedding function. To address this issue, we propose a mechanism that leverages the statistics of the nearest neighbors of a sample. Our approach utilizes cross-attention to learn meaningful information from other samples, thereby improving the local embedding of the image. This method can be easily incorporated into DML approaches at a negligible additional cost during inference. Experimental results on various standard DML benchmark datasets demonstrate that our approach outperforms the state of the art.

Primary Area: metric learning, kernel learning, and sparse coding

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7269