Abstract
Low-resolution (LR) face recognition poses a significant challenge in embedding learning due to the severe loss of identity information. Recent softmax losses introduce a non-static margin that gives greater importance or a larger margin to recognizable examples based on the embedding norm as a measure of face recognizability. In this paper, we argue that face recognizability is more than just the embedding norm, as it does not capture the fine-grained details of face images that are critical to embedding learning. We propose SlackedFace to induce a relaxed margin aligned with face recognizability and the model's confidence based on both embedding norm and embedding proximity for empowered embedding learning. We also introduce fast-hill climbing as an early calibration stage between pre-trained and randomly initialized modules. We show that SlackedFace outperforms the current best models on realistic LR face datasets when tested in practical open-set evaluation scenarios.Video
Citation
@inproceedings{Low_2023_BMVC,
author = {Cheng Yaw Low and Jacky Chen Long Chai and Jaewoo Park and KYEONGJIN ANN and Meeyoung Cha},
title = {SlackedFace: Learning a Slacked Margin for Low-Resolution Face Recognition},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0282.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by