Improving Deep Metric Learning by Divide and Conquer.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep metric learning (DML) maps data to an embedding space for similarity-based tasks.
• Current DML methods struggle to generalize to unseen categories due to varied visual factors.
• Existing approaches often learn a single embedding space, failing to capture diverse object relationships.

Purpose of the Study:

• To develop a more expressive representation for improved DML generalization.
• To address the limitations of single embedding spaces in capturing multifaceted object similarities.
• To enhance performance on tasks like image retrieval, clustering, and re-identification.

Main Methods:

• Proposes a hierarchical splitting of the embedding space and data into smaller subsets.
• Learns separate embedding subspaces for each data subset, reducing variance.
• Jointly learns subspaces to encompass both intricate details and broad data characteristics.
• Combines subspaces in a 'conquering' stage to form the final embedding.
• Acts as a transparent wrapper applicable to existing DML methods.

Main Results:

• Significantly improves state-of-the-art performance on image retrieval, clustering, and re-identification.
• Demonstrates superior generalization capabilities on novel categories.
• Validated across diverse datasets including CUB200-2011, CARS196, Stanford Online Products, In-shop Clothes, and PKU VehicleID.

Conclusions:

• The proposed hierarchical DML approach offers enhanced representational power and generalization.
• This method effectively encodes diverse visual relationships, overcoming limitations of single embedding spaces.
• The algorithm provides a versatile enhancement for various DML techniques and downstream applications.