Pairwise Similarity Learning is SimPLE
Yandong Wen*, Weiyang Liu*, Yao Feng, Bhiksha Raj, Rita Singh,Adrian Weller, Michael J. Black, Bernhard Schölkopf
*authors contributed equally
ICCV 2023, Paris, France
[Paper] [Code]
Abstract
In this paper, we focus on a general yet important learning problem, pairwise similarity learning (PSL). PSL subsumes a wide range of important applications, such as open-set face recognition, speaker verification, image retrieval and person re-identification. The goal of PSL is to learn a pairwise similarity function assigning a higher similarity score to positive pairs (i.e., a pair of samples with the same label) than to negative pairs (i.e., a pair of samples with different label). We start by identifying a key desideratum for PSL, and then discuss how existing methods can achieve this desideratum. We then propose a surprisingly simple proxy-free method, called SimPLE, which requires neither feature/proxy normalization nor angular margin and yet is able to generalize well in open-set recognition. We apply the proposed method to three challenging PSL tasks: open-set face recognition, image retrieval and speaker verification. Comprehensive experimental results on large-scale benchmarks show that our method performs significantly better than current state-of-the-art methods.
Results on IJB-B and IJB-C
Method | IJB-B | IJB-C | ||||||||||
1:1 Verification Rate TPR@FPR | 1:N Identification Rate TPIR@FPIR | 1:1 Verification Rate TPR@FPR | 1:N Identification Rate TPIR@FPIR | |||||||||
1e-6 | 1e-5 | 1e-4 | top 1 | 1e-2 | 1e-1 | 1e-6 | 1e-5 | 1e-4 | top 1 | 1e-2 | 1e-1 | |
SphereFace | 47.33 | 90.14 | 94.87 | 95.13 | 82.57 | 94.30 | 87.86 | 94.36 | 96.25 | 96.45 | 91.68 | 95.36 |
CosFace | 43.67 | 88.83 | 95.23 | 95.35 | 80.50 | 94.49 | 85.29 | 94.33 | 96.62 | 96.53 | 90.69 | 95.61 |
ArcFace | 43.43 | 90.40 | 95.02 | 95.14 | 81.36 | 94.26 | 86.00 | 94.49 | 96.39 | 96.47 | 91.91 | 95.51 |
CurricularFace† | - | - | 94.86 | - | - | - | - | - | 96.15 | - | - | - |
BroadFace† | 40.92 | 89.97 | 94.97 | - | - | - | 85.96 | 94.59 | 96.38 | - | - | - |
SCF-ArcFace† | - | 90.68 | 94.74 | - | - | - | - | 94.04 | 96.09 | - | - | - |
SphereFace2 | 41.53 | 89.92 | 95.02 | 95.24 | 83.46 | 94.36 | 87.63 | 94.49 | 96.42 | 96.41 | 92.08 | 95.47 |
MagFace+ | 42.32 | 90.36 | 94.51 | 94.81 | 83.65 | 93.87 | 90.24 | 94.08 | 95.97 | 96.02 | 91.95 | 95.06 |
AdaFace | 46.78 | 90.04 | 95.67 | 95.54 | 80.73 | 95.07 | 89.74 | 94.87 | 96.89 | 96.75 | 92.12 | 96.20 |
SimPLE | 49.87 | 91.13 | 94.78 | 95.54 | 85.92 | 94.28 | 90.30 | 94.34 | 96.27 | 96.81 | 92.88 | 95.49 |
† Results are obtained from their papers.
Acknowledgement
The authors would like to thank colleges from Max Planck Institute for Intelligent Systems at Tübingen for many inspiring discussions. This work was supported by the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039B, and by the Machine Learning Cluster of Excellence, EXC number 2064/1 – Project number 390727645.
WL was supported by the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, TP XX, project number: 276693517.
AW acknowledges support from a Turing AI Fellowship under grant EP/V025379/1, and the Leverhulme Trust via Leverhulme Centre for the Future of Intelligence.
MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon.MJB has financial interests in Amazon, Datagen Technologies, and Meshcapade GmbH. While MJB is a consultant for Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.
Reference
@inproceedings{wen2023simple,
author = {Yandong Wen*, Weiyang Liu*, Yao Feng, Bhiksha Raj, Rita Singh, Adrian Weller, Michael J. Black, Bernhard Sch\"olkopf},
title = {Pairwise Similarity Learning is SimPLE},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
}