MLFW (Masked LFW)

由于当前的 COVID-19 大流行，随着越来越多的人开始戴口罩，现有的人脸识别系统在识别蒙面人脸时可能会遇到严重的性能下降。为了弄清楚口罩对人脸识别模型的影响，我们构建了一个简单但有效的工具来自动从未蒙面的人脸生成蒙面人脸，并基于跨年龄 LFW (CALFW) 数据库构建了一个名为 Masked LFW (MLFW) 的新数据库。我们的方法生成的蒙面脸上的蒙版与原始人脸具有良好的视觉一致性。此外，我们收集了各种面具模板，涵盖了日常生活中出现的大部分常见样式，以实现多样化的生成效果。考虑到现实场景，我们设计了三种人脸对组合。 SOTA模型在MLFW数据库上的识别准确率比在原始图像上的准确率下降了5%-16%。构建 MLFW 基准的背后有以下三个动机：
1.建立一个相对难度更大的数据库来评估蒙面人脸验证的性能，从而充分证明几种人脸验证方法的有效性。
2.超越年龄差距，MLFW认为两个身份相同的人脸戴不同的面具，两个不同身份的人脸戴相同的面具，这进一步强调了同时大的类内方差和微小的类间方差。
3.保持数据大小，提供“相同/不同”基准的人脸验证协议，同时保持与calfw中相同的身份。因此，MLFW 可以很容易地应用于评估蒙面人脸验证的性能。

Amid the ongoing COVID-19 pandemic, as an increasing number of people start wearing face masks, existing face recognition systems may suffer from severe performance degradation when identifying masked faces. To clarify the impact of face masks on face recognition models, we constructed a simple yet effective tool to automatically generate masked faces from unmasked facial images, and built a new database named Masked LFW (MLFW) based on the cross-age LFW (CALFW) dataset. The masks generated by our method exhibit strong visual consistency with the original faces. In addition, we collected various mask templates covering most common styles appearing in daily life to achieve diverse generation results. Considering real-world scenarios, we designed three types of face pair combinations. The recognition accuracy of state-of-the-art (SOTA) models on the MLFW database decreased by 5%-16% compared to that on original images. There are three core motivations behind the construction of the MLFW benchmark:
1. Establish a relatively more challenging database to evaluate the performance of masked face verification, thereby fully demonstrating the effectiveness of several face verification methods.
2. Beyond age gaps, MLFW considers scenarios where two faces of the same identity wear different masks, and two faces of different identities wear the same mask, which further emphasizes both large intra-class variance and small inter-class variance.
3. Maintain the dataset size and provide the "same/different" face verification protocol while keeping the identical identities as in CALFW. Thus, MLFW can be easily applied to evaluate the performance of masked face verification.

搜集汇总

背景与挑战

背景概述

MLFW是一个用于评估蒙面人脸识别性能的数据集，通过自动生成蒙面人脸并设计多种人脸对组合，模拟现实场景中的戴口罩情况。该数据集旨在挑战现有模型的性能，并提供一个标准化的评估基准。