Optimizing Rank-based Metrics with Blackbox Differentiation
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.
| Author(s): | Michal Rolínek and Vít Musil and Anselm Paulus and Marin Vlastelica and Claudio Michaelis and Georg Martius |
| Book Title: | 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020) |
| Pages: | 7617 -- 7627 |
| Year: | 2020 |
| Month: | June |
| Day: | 14-19 |
| Publisher: | IEEE |
| Project(s): | |
| Bibtex Type: | Conference Paper (inproceedings) |
| Address: | Piscataway, NJ |
| DOI: | 10.1109/CVPR42600.2020.00764 |
| Event Name: | IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR 2020) |
| Event Place: | Seattle, WA, USA |
| State: | Published |
| URL: | https://openaccess.thecvf.com/content_CVPR_2020/html/Rolinek_Optimizing_Rank-Based_Metrics_With_Blackbox_Differentiation_CVPR_2020_paper.html |
| Digital: | True |
| Electronic Archiving: | grant_archive |
| ISBN: | 978-1-7281-7168-5 |
| Note: | Best paper nomination |
| Talk Type: | Oral |
| Links: | |
BibTex
@inproceedings{Rolinek2020optimizing,
title = {Optimizing Rank-based Metrics with Blackbox Differentiation},
booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020)},
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. },
pages = {7617 -- 7627},
publisher = {IEEE},
address = {Piscataway, NJ},
month = jun,
year = {2020},
note = {Best paper nomination},
slug = {optimizing-rank-based-metrics-with-blackbox-differentiation},
author = {Rolínek, Michal and Musil, Vít and Paulus, Anselm and Vlastelica, Marin and Michaelis, Claudio and Martius, Georg},
url = {https://openaccess.thecvf.com/content_CVPR_2020/html/Rolinek_Optimizing_Rank-Based_Metrics_With_Blackbox_Differentiation_CVPR_2020_paper.html},
month_numeric = {6}
}