MFR (Ongoing version of ICCV-2021 Masked Face Recognition Challenge & Workshop(MFR))

在 COVID-19 冠状病毒流行期间，几乎每个人都戴着口罩，这对人脸识别提出了巨大挑战。传统的人脸识别系统可能无法有效识别蒙面人脸，但摘下口罩进行身份验证会增加病毒感染的风险。受 COVID-19 大流行应对措施的启发，人们在公共场所佩戴防护口罩的广泛要求促使人们需要了解人脸识别技术如何处理被遮挡的面部，通常只有眼周区域及以上可见。为了应对戴口罩带来的挑战，改进现有的人脸识别方法至关重要。最近，一些商业供应商宣布推出能够处理口罩的人脸识别算法，越来越多的研究出版物出现在戴口罩的人脸识别主题上。但是，由于疫情的突然爆发，目前还没有公开的蒙面人脸识别基准。在本次工作坊中，我们将组织蒙面人脸识别（MFR）挑战赛，重点关注在口罩存在下对深度人脸识别方法进行基准测试。在这个挑战中，我们将评估以下测试集的准确性： 蒙面和非蒙面人脸之间的准确性。儿童（2~16岁）的准确性。全球化多种族基准的准确性。我们确保这些测试集和公开可用的训练数据集之间没有重叠，因为它们不是从在线名人那里收集的。全球化的多种族测试集包含 242,143 个身份和 1,624,305 个图像。掩码测试集包含 6,964 个身份、6,964 个掩码图像和 13,928 个非掩码图像。总共有 13,928 个阳性对和 96,983,824 个阴性对。儿童测试集包含 14,344 个身份和 157,280 张图像。总共有 1,773,428 个阳性对和 24,735,067,692 个阴性对。对于 Mask 集，TAR 是在 mask-to-non-mask 1:1 协议上测量的，FAR 小于 0.0001(e-4)。对于儿童组，TAR 是在全对全 1:1 协议上测量的，FAR 小于 0.0001(e-4)。对于其他集合，TAR 是在全对全 1:1 协议上测量的，FAR 小于 0.000001(e-6)。参与者按两个数据集的最高分数排序：TAR@Mask 和 TAR@MR-All，公式为 0.25 * TAR@Mask + 0.75 * TAR@MR-All。

During the COVID-19 pandemic, nearly everyone wore face masks, which posed significant challenges to facial recognition systems. Traditional facial recognition systems may fail to effectively recognize masked faces, while removing masks for identity verification would increase the risk of viral infection. Inspired by the COVID-19 pandemic response measures, the widespread requirement of wearing protective masks in public spaces has driven the need to understand how facial recognition technologies handle occluded faces, where typically only the periorbital region and above are visible. To address the challenges brought by mask-wearing, improving existing facial recognition methods is critical. Recently, several commercial vendors have announced facial recognition algorithms capable of handling masked faces, and an increasing number of research publications have emerged on the topic of masked facial recognition. However, due to the sudden outbreak of the pandemic, there are currently no publicly available masked facial recognition benchmarks.

In this workshop, we will organize a Masked Facial Recognition (MFR) challenge focusing on benchmarking deep facial recognition methods in the presence of masks. In this challenge, we will evaluate accuracy across the following test sets: accuracy between masked and unmasked faces; accuracy for children aged 2 to 16; and accuracy for the global multi-racial benchmark. We ensure there is no overlap between these test sets and publicly available training datasets, as they are not collected from online celebrities.

The global multi-racial test set contains 242,143 identities and 1,624,305 images. The mask test set includes 6,964 identities, 6,964 masked images, and 13,928 unmasked images. There are a total of 13,928 positive pairs and 96,983,824 negative pairs. The child test set contains 14,344 identities and 157,280 images. There are a total of 1,773,428 positive pairs and 24,735,067,692 negative pairs.

For the Mask set, TAR is measured under the mask-to-non-mask 1:1 protocol, with FAR less than 0.0001 (1e-4). For the child group, TAR is measured under the all-to-all 1:1 protocol, with FAR less than 0.0001 (1e-4). For other sets, TAR is measured under the all-to-all 1:1 protocol, with FAR less than 0.000001 (1e-6). Participants are ranked based on the weighted composite score of two datasets: TAR@Mask and TAR@MR-All, calculated using the formula 0.25 * TAR@Mask + 0.75 * TAR@MR-All.

搜集汇总

背景与挑战

背景概述

MFR数据集是一个专注于佩戴口罩情况下人脸识别技术性能评估的基准数据集，包含多种测试集以评估不同场景下的识别准确性，特别关注儿童和多种族情况。该数据集由清华大学和伦敦帝国理工学院于2021年发布，旨在推动口罩遮挡下的人脸识别技术发展。