halfFace: A face covering mask dataset of South Asian people
收藏资源简介：
halfFace is a dataset that contains facial images of both masked and unmasked images of South Asian people (166 participants). This is a combination of 3 different datasets which include full face, upper face, and augmented upper face images so that it can be used for person identification from both masked and unmasked images and detecting unmasked person. The participants were fully informed about the usage of the dataset and an approval was taken from them for further usage of this data while collecting data from them. Method Images of the dataset were captured in participant’s environment, and using participant’s smart phone camera. Hence, this dataset contains the diversity of devices. This images were then collected by the authors. Then, three different datasets were developed from the images. In the dataset-1, only facial part was kept using a pretrained CNN model, and it was not divided into train-test images. The facial images of masked and unmasked images can be easily identified by naming convention. In dataset-2, we have excluded the covered face portion by using YOLOV3. So, images of dataset-2 contains only upper part of the face (hair, forehead, eyes, and upper portion of nose), and it was further divided into training-testing parts. As models require a huge amount of data to train, we have increased the number of training images using augmentation method by adjusting brightness, blur, contrast, saturation, gaussian noise, and salt and pepper noise. This has made the dataset-3 even more diverse and noisy which is essential in building robust models. Usage Notes Dataset-1 : Full face images Datset-2: Upper face without augmentation (divided into train-test) Dataset-3: Upper face with augmentation of training part, testing images were same as dataset-2 Naming convention: Folders: P_XXX where XXX is the identification of participants. Images of dataset-1 and dataset-2: pXXX_(0/1)(0/1/2)(m/f) Here, pXXX is the person identification number. First (0/1) means unmasked (0) or masked(1). (0/1/2) means the image is front facing(0)/ left facing (1), right facing (2). (m/f) denotes the gender of the participant. Images of dataset-3: pXXX_(0/1)(0/1/2)(m/f)_(b/blur/c/GN/s/SPN) Here, pXXX is the person identification number. First (0/1) means unmasked (0) or masked(1). (0/1/2) means the image is front facing(0)/ left facing (1), right facing (2). (m/f) denotes the gender of the participant. “b/blur/c/GN/s/SPN” means change of brightness (b), blur (blur), contrast (c), gaussian noise (GN), saturation (s), Salt and Pepper noise (SPN).
halfFace 是一个涵盖166名南亚参与者戴口罩与未戴口罩面部图像的数据集。该数据集整合了3个子数据集，分别包含全脸图像、上脸图像及经过数据增强的上脸图像，可用于基于戴口罩与未戴口罩图像的个体身份识别，以及未戴口罩人员检测。 数据采集前，所有参与者已被充分告知数据集的使用用途，并征得其同意后方可开展后续的数据采集与复用工作。 数据集的图像均在参与者所处的环境中，使用参与者自身的智能手机摄像头拍摄，因此该数据集涵盖了多样化的采集设备。所有图像均由数据集开发者收集。随后基于原始图像构建了3个子数据集： 1. 数据集1：通过预训练卷积神经网络（Convolutional Neural Network, CNN）仅保留面部区域，未划分训练集与测试集；戴口罩与未戴口罩的面部图像可通过命名规则直接区分。 2. 数据集2：通过YOLOV3算法去除了被口罩遮挡的面部区域，因此仅保留面部上半部分（包括头发、前额、眼部及鼻上部），并进一步划分为训练集与测试集。 3. 考虑到模型训练需要大规模数据支撑，我们通过数据增强手段扩充训练集图像数量，具体操作包括调整亮度、模糊度、对比度、饱和度，添加高斯噪声与椒盐噪声。由此构建的数据集3进一步提升了数据多样性与噪声复杂度，这对构建鲁棒性模型至关重要。 使用说明 - 数据集1：全脸图像 - 数据集2：未经过数据增强的上脸图像（已划分为训练集与测试集） - 数据集3：训练集经过数据增强的上脸图像，测试集与数据集2完全一致 命名规则 文件夹命名格式：P_XXX，其中XXX为参与者的标识编号。 数据集1与数据集2的图像命名格式：pXXX_(0/1)(0/1/2)(m/f) 其中，pXXX为参与者编号；(0/1)分别代表未戴口罩（0）与戴口罩（1）；(0/1/2)分别代表正面朝向（0）、左侧朝向（1）与右侧朝向（2）；(m/f)代表参与者性别。 数据集3的图像命名格式：pXXX_(0/1)(0/1/2)(m/f)_(b/blur/c/GN/s/SPN) 其中，pXXX为参与者编号；(0/1)分别代表未戴口罩（0）与戴口罩（1）；(0/1/2)分别代表正面朝向（0）、左侧朝向（1）与右侧朝向（2）；(m/f)代表参与者性别；"b/blur/c/GN/s/SPN"分别代表亮度调整（b）、模糊处理（blur）、对比度调整（c）、高斯噪声（Gaussian Noise, GN）、饱和度调整（s）与椒盐噪声（Salt and Pepper Noise, SPN）。
