Optimizing Rank-based Metrics with Blackbox Differentiation

D 2020

Optimizing Rank-based Metrics with Blackbox Differentiation

ROLINEK, Michal; Vít MUSIL; Anselm PAULUS; Marin VLASTELICA; Claudio MICHAELIS et al.

Základní údaje

Originální název

Optimizing Rank-based Metrics with Blackbox Differentiation

Autoři

ROLINEK, Michal; Vít MUSIL; Anselm PAULUS; Marin VLASTELICA; Claudio MICHAELIS a Georg MARTIUS

Vydání

Los Alamitos, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020), od s. 7617-7627, 11 s. 2020

Nakladatel

IEEE COMPUTER SOC

Další údaje

Jazyk

angličtina

Typ výsledku

Stať ve sborníku

Obor

10101 Pure mathematics

Utajení

není předmětem státního či obchodního tajemství

Forma vydání

elektronická verze "online"

Odkazy

Springer

Organizační jednotka

Fakulta informatiky

ISSN

DOI

https://doi.org/10.1109/CVPR42600.2020.00764

UT WoS

001309199900027

EID Scopus

2-s2.0-85094614729

Příznaky

Mezinárodní význam, Recenzováno

Změněno: 7. 11. 2024 10:23, RNDr. Vít Musil, Ph.D.

Anotace

V originále

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.

Citovat

ROLINEK, Michal; Vít MUSIL; Anselm PAULUS; Marin VLASTELICA; Claudio MICHAELIS a Georg MARTIUS. Optimizing Rank-based Metrics with Blackbox Differentiation. Online. In 978-1-7281-7168-5. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020). Los Alamitos: IEEE COMPUTER SOC, 2020, s. 7617-7627. ISSN 1063-6919. Dostupné z: https://doi.org/10.1109/CVPR42600.2020.00764.

@inproceedings{2449799,
   author = {Rolinek, Michal and Musil, Vít and Paulus, Anselm and Vlastelica, Marin and Michaelis, Claudio and Martius, Georg},
   address = {Los Alamitos},
   booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020)},
   doi = {https://doi.org/10.1109/CVPR42600.2020.00764},
   editor = {978-1-7281-7168-5},
   howpublished = {elektronická verze "online"},
   language = {eng},
   location = {Los Alamitos},
   pages = {7617-7627},
   publisher = {IEEE COMPUTER SOC},
   title = {Optimizing Rank-based Metrics with Blackbox Differentiation},
   url = {https://link.springer.com/chapter/10.1007/978-3-030-58604-1_25},
   year = {2020}
}

TY  - CONF
ID  - 2449799
AU  - Rolinek, Michal - Musil, Vít - Paulus, Anselm - Vlastelica, Marin - Michaelis, Claudio - Martius, Georg
PY  - 2020
TI  - Optimizing Rank-based Metrics with Blackbox Differentiation
PB  - IEEE COMPUTER SOC
CY  - Los Alamitos
UR  - https://link.springer.com/chapter/10.1007/978-3-030-58604-1_25
N2  - 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.
ER  -

ROLINEK, Michal; Vít MUSIL; Anselm PAULUS; Marin VLASTELICA; Claudio MICHAELIS a Georg MARTIUS. Optimizing Rank-based Metrics with Blackbox Differentiation. Online. In 978-1-7281-7168-5. \textit{2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020)}. Los Alamitos: IEEE COMPUTER SOC, 2020, s.~7617-7627. ISSN~1063-6919. Dostupné z: https://doi.org/10.1109/CVPR42600.2020.00764.

Přehled o publikaci