CelebA+masks

COVID-19 大流行引发了使人脸识别系统适应新现实的问题，人们可能会戴上外科口罩来遮住鼻子和嘴巴。用于训练这些系统的传统数据集（例如 CelebA、CASIA-WebFace）是在大流行之前发布的，因此由于缺乏戴口罩的人的例子，它们现在似乎不适合。我们提出了一种通过创建合成蒙版并将它们覆盖在原始图像中的面部上来增强包含没有蒙版的人脸的数据集的方法。我们的方法依赖于 Spark AR Studio，这是 Facebook 开发的一个开发程序，用于创建 Instagram 人脸过滤器。在我们的方法中，我们使用了 9 个不同颜色、形状和面料的面具。我们使用我们的方法为 CelebA 数据集生成了 196,254 个 (96.8%) 掩码。

COVID-19 pandemic has posed a critical challenge for adapting facial recognition systems to new real-world scenarios, where people may wear surgical masks to cover their noses and mouths. Traditional datasets used for training these systems (e.g., CelebA, CASIA-WebFace) were released prior to the COVID-19 pandemic, and thus they are now unsuitable due to the absence of samples of individuals wearing face masks. We propose a method to enhance datasets containing unmasked facial images by generating synthetic masks and overlaying them onto the facial regions in the original images. Our method leverages Spark AR Studio, a development program developed by Facebook for creating Instagram facial filters. In our approach, we employ 9 distinct masks with varying colors, shapes and fabrics. We applied our method to generate 196,254 (96.8%) masked facial images for the CelebA dataset.

搜集汇总

背景与挑战

背景概述

CelebA+masks是一个增强版的人脸数据集，通过在CelebA原始图像上合成9种不同口罩，生成了196,254个带口罩的人脸图像，旨在解决COVID-19大流行期间戴口罩人脸识别的需求。该数据集由布加勒斯特大学于2021年发布，适用于人脸识别研究。