Optimization of Rank Losses for Image Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Standard image retrieval evaluation relies on ranking metrics like average precision (AP), recall at k (R@k), and normalized discounted cumulative gain (NDCG).
• End-to-end training of deep neural networks with rank losses faces challenges due to non-differentiability and non-decomposability.

Purpose of the Study:

• To develop a general framework for robust and decomposable rank losses optimization in deep neural networks.
• To address the limitations of current methods in training models with ranking-based evaluation metrics.

Main Methods:

• Proposed a general surrogate for the ranking operator, SupRank, which is amenable to stochastic gradient descent and provides an upper bound for rank losses.
• Introduced a loss function to reduce the decomposability gap between batch approximations and full training set values of rank losses.
• Extended the framework to hierarchical image retrieval, introducing hierarchical average precision (H-AP).

Main Results:

• The SupRank surrogate enables robust training for deep neural networks using rank losses.
• The proposed loss function effectively bridges the gap in decomposability for rank losses.
• The framework successfully applied to AP and R@k metrics, and extended for hierarchical image retrieval.
• Developed the first hierarchical landmarks retrieval dataset using a semi-automatic pipeline on Google Landmarks v2.

Conclusions:

• The developed framework offers a robust and decomposable approach to optimizing rank losses for deep learning models in image retrieval.
• The framework's applicability to standard and hierarchical metrics, along with the creation of a new dataset, advances the field of image retrieval research.