Optimizing Rank-based Metrics with Blackbox Differentiation

TL;DR

This paper introduces a novel blackbox differentiation method for directly optimizing non-differentiable rank-based metrics like recall and Average Precision in computer vision tasks, improving performance on image retrieval and object detection.

Contribution

It provides a theoretically sound, efficient approach for differentiating rank-based metrics using mini-batch gradient descent, addressing optimization challenges such as instability and sparsity.

Findings

Achieves competitive performance on image retrieval datasets.

Improves performance of near state-of-the-art object detectors.

Provides a general method applicable to various rank-based metrics.

Abstract

Rank-based metrics are some of the most widely used criteria for performance evaluation of computer vision models. Despite years of effort, direct optimization for these metrics remains a challenge due to their non-differentiable and non-decomposable nature. We present an efficient, theoretically sound, and general method for differentiating rank-based metrics with mini-batch gradient descent. In addition, we address optimization instability and sparsity of the supervision signal that both arise from using rank-based metrics as optimization targets. Resulting losses based on recall and Average Precision are applied to image retrieval and object detection tasks. We obtain performance that is competitive with state-of-the-art on standard image retrieval datasets and consistently improve performance of near state-of-the-art object detectors. The code is available at https://github.com/martius-lab/blackbox-backprop